Risk layering and decision-making method based on interaction of large language model and risk model

Abstract

The invention discloses a risk layering and decision-making method based on interaction of a large language model and a risk model, and the method comprises the steps: obtaining and processing multi-modal clinical data of a patient, generating a structured feature matrix, and inputting a risk prediction model to obtain a multi-dimensional risk spectrum; retrieving evidence-based evidence from the medical knowledge base based on the risk probability exceeding the threshold; combining the feature matrix, the risk spectrum and the evidence, and generating an explanatory report through a large language model and cue word engineering; the model weight or retrieval strategy is dynamically adjusted by comparing the consistency of the intervention priority and the risk probability; performing risk layering according to the adjusted risk spectrum and generating personalized intervention suggestions; and continuously optimizing the model according to a medical intervention result and patient feedback data to form closed-loop iterative updating. According to the method, dynamic risk prediction, explanatory report generation and personalized intervention suggestion are realized, and meanwhile, the clinical application value and decision accuracy are improved based on a doctor feedback and patient data closed-loop optimization model.

Description

Risk Stratification and Decision-Making Method Based on the Interaction of Large Language Model and Risk Model Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a risk stratification and decision-making method based on the interaction of a large language model and a risk model. Background Technology

[0002] In recent years, with the development of medical big data and artificial intelligence, disease risk prediction and personalized intervention have gradually become important directions in precision medicine. Traditional risk prediction models (such as the Cox proportional hazards model, logistic regression, and multilayer perceptron) have been widely used in chronic disease management, cardiovascular disease risk assessment, and tumor prognosis prediction. These models typically rely on a large amount of historical data for training, and by inputting variables such as clinical indicators, laboratory tests, and imaging data, they output individual risk scores or stratification levels, providing doctors with risk references.

[0003] Meanwhile, Large Language Models (LLMs) have shown great potential in medical knowledge retrieval, clinical guideline interpretation, and doctor-patient interaction. Through techniques such as Retrieval-Augmented Generation (RAG) and cue word engineering, LLMs can establish connections between unstructured medical knowledge and structured clinical data, generating interpretable reports or auxiliary treatment suggestions. Existing research has achieved breakthroughs in electronic medical record summary generation, evidence-based medicine-assisted retrieval, and intelligent question answering.

[0004] Some existing research has attempted to combine risk prediction models with intelligent interactive systems. For example, in kidney disease management, some scholars have proposed combining electronic medical records with risk prediction models to assist in assessing a patient's risk of cardiovascular complications. In chronic disease management platforms, some systems can push corresponding health advice based on simple parameter thresholds. However, overall, the technological integration of these methods is limited, mainly confined to a one-way information flow, i.e., "input data—risk model calculation—output score," lacking deep interaction and multi-dimensional interpretation. Specifically, current technologies mainly suffer from the following shortcomings: 1) Risk prediction results lack interpretability and traceability. Traditional models often output a risk score or risk level, but struggle to provide multi-dimensional interpretations. For example, a patient may be classified as "high-risk," but clinicians cannot clearly understand which indicators drove this result, nor can they quickly correlate it with evidence-based medicine. This makes doctors lack confidence in risk management and makes it difficult to directly translate the results into intervention measures.

[0005] 2) There is a disconnect between the model and the knowledge base. Most risk prediction models rely on structured data for reasoning, while clinical decision-making often requires the integration of unstructured medical knowledge (such as guidelines, meta-analyses, and RCT evidence). Existing methods have not yet established a mechanism that allows risk prediction results to be effectively integrated with the medical knowledge base to form a closed-loop chain of evidence.

[0006] 3) Inadequate individualized intervention recommendations. Some systems can provide simple "follow-up recommendations" or "medication reminders" after identifying high-risk patients. However, for complex patients (such as those with overlapping risks of multiple organ complications), there is a lack of refined decision support such as tiered intervention strategies, multi-organ protection pathways, and dynamic follow-up adjustments.

[0007] 4) Lack of dynamic updates and closed-loop optimization. Most risk models lack a real-time update mechanism after deployment. Clinical data is often dynamic, such as blood pressure, inflammatory markers, and dialysis parameters. If the model cannot dynamically adjust for risk based on new data, it will be difficult to achieve precise management. At the same time, the results of physician interventions and patient feedback fail to feed back into the model, making it difficult for the system to be continuously optimized.

[0008] 5) Insufficient clinical application value. Although some studies have achieved risk prediction, in actual clinical applications, doctors often face problems such as complex operation, difficulty in interpreting results, and generalization of intervention recommendations, making it difficult to implement the system. Especially in high-risk areas such as kidney disease and cardiovascular disease, how to achieve truly usable decision support through intelligent interactive systems remains an unsolved problem.

[0009] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention provides a risk stratification and decision-making method, system, electronic device, and storage medium based on the interaction of a large language model and a risk model.

[0011] In a first aspect, the present invention provides a risk stratification and decision-making method based on the interaction of a large language model and a risk model. The technical solution of this method is as follows: Acquire multimodal clinical data of patients, and preprocess and extract features from the multimodal clinical data to obtain a structured feature matrix; input the structured feature matrix into a pre-trained risk prediction model to output a multidimensional risk spectrum, wherein the multidimensional risk spectrum includes the patient's risk probability on multiple clinical outcomes; based on the risk probability exceeding a preset threshold in the multidimensional risk spectrum, retrieve corresponding evidence-based evidence from a medical knowledge base; input the structured feature matrix, the multidimensional risk spectrum, and the evidence-based evidence into the large language model. An explanatory report is generated through prompt word engineering; the intervention priority for each clinical outcome in the explanatory report is compared with the risk probability of the corresponding clinical outcome in the multidimensional risk spectrum; if the intervention priority for the same clinical outcome is inconsistent with the risk level indicated by the risk probability, the feature weights of the risk prediction model are dynamically adjusted or the evidence retrieval strategy is updated; the patient is risk-stratified according to the dynamically adjusted multidimensional risk spectrum and the preset stratification threshold, and personalized intervention suggestions are generated; based on the intervention measures implemented by the doctor and the patient feedback data, the risk prediction model and the large language model are dynamically optimized to achieve closed-loop iterative updates.

[0012] The beneficial effects of the risk stratification and decision-making method based on the interaction of a large language model and a risk model of the present invention are as follows: The method of the present invention can solve the problems of insufficient interpretability and separation from the knowledge base of traditional risk prediction models through the interaction of a large language model and a risk model, and realizes dynamic risk prediction, interpretive report generation and personalized intervention suggestions. At the same time, based on the closed-loop optimization of the model based on doctor feedback and patient data, it improves the clinical application value and decision-making accuracy.

[0013] Based on the above scheme, the risk stratification and decision-making method of the present invention based on the interaction of a large language model and a risk model can be further improved as follows.

[0014] In one optional approach, the steps of preprocessing and feature extraction of the clinical data to obtain a structured feature matrix include: processing missing values ​​in the clinical data using multiple imputation or KNN imputation methods, and removing or correcting outliers in the clinical data based on statistical distribution or medical thresholds to obtain preprocessed clinical data; and performing feature filtering on the preprocessed clinical data using principal component analysis or LASSO methods to obtain the structured feature matrix.

[0015] In one alternative approach, the multidimensional risk spectrum is also visualized using a radar chart or heat map.

[0016] In one alternative approach, the step of retrieving corresponding evidence-based evidence from a medical knowledge base includes: retrieving evidence-based evidence corresponding to each clinical outcome for which the risk probability exceeds a preset threshold from the medical knowledge base using a vector retrieval method based on the FAISS index.

[0017] In one alternative approach, the medical knowledge base includes at least one of evidence-based medicine guidelines, meta-analyses, randomized controlled trial literature, and professional databases.

[0018] In one alternative approach, the step of inputting the structured feature matrix, the multidimensional risk spectrum, and the evidence-based evidence into a large language model and generating an interpretive report through cue word engineering includes: constructing a target instruction through the cue word engineering to cause the large language model to execute the target instruction; analyzing the correlation between the risk probability in the multidimensional risk spectrum and the evidence-based evidence; and generating the interpretive report, which includes attribution of risk drivers, explanation of evidence source, and intervention priority for each clinical outcome.

[0019] In one alternative approach, the step of risk stratifying the patient based on a dynamically adjusted multidimensional risk spectrum and a preset stratification threshold includes: comparing the risk probability of each clinical outcome in the dynamically adjusted multidimensional risk spectrum with the preset stratification threshold, and classifying the patient into a low-risk group, a medium-risk group, or a high-risk group.

[0020] Secondly, this invention provides a risk stratification and decision-making system based on the interaction of a large language model and a risk model. The technical solution of this system is as follows: a data acquisition module, used to acquire multimodal clinical data of patients, and preprocess and extract features from the multimodal clinical data to obtain a structured feature matrix; a risk prediction module, used to input the structured feature matrix into a pre-trained risk prediction model, outputting a multidimensional risk spectrum, the multidimensional risk spectrum including the patient's risk probability on multiple clinical outcomes; a knowledge enhancement module, used to retrieve corresponding evidence-based evidence from a medical knowledge base based on the risk probability exceeding a preset threshold in the multidimensional risk spectrum; and a model interaction module, used to input the structured feature matrix, the multidimensional risk spectrum, and the evidence-based evidence into the large language model. In the language model, an explanatory report is generated through prompt word engineering; a dynamic adjustment module is used to compare the intervention priority for each clinical outcome in the explanatory report with the risk probability of the corresponding clinical outcome in the multi-dimensional risk spectrum; if the intervention priority for the same clinical outcome is inconsistent with the risk level indicated by the risk probability, the feature weights of the risk prediction model are dynamically adjusted or the evidence retrieval strategy is updated; a risk stratification module is used to stratify the patient's risk based on the dynamically adjusted multi-dimensional risk spectrum and a preset stratification threshold, and generate personalized intervention suggestions; a closed-loop feedback module is used to dynamically optimize the risk prediction model and the large language model based on the intervention measures implemented by the doctor and the patient feedback data, so as to achieve closed-loop iterative updates.

[0021] The beneficial effects of the risk stratification and decision-making system based on the interaction of a large language model and a risk model of the present invention are as follows: The system of the present invention can solve the problems of insufficient interpretability and separation from the knowledge base of traditional risk prediction models through the interaction of a large language model and a risk model, and realizes dynamic risk prediction, interpretive report generation and personalized intervention suggestions. At the same time, based on the closed-loop optimization of the model with doctor feedback and patient data, it improves the clinical application value and decision-making accuracy.

[0022] Thirdly, the technical solution of an electronic device according to the present invention is as follows: it includes a memory, a processor, and a program stored in the memory and running on the processor, wherein when the processor executes the program, it implements the steps of the risk stratification and decision-making method based on the interaction of a large language model and a risk model as described in the present invention.

[0023] Fourthly, the present invention provides a computer-readable storage medium with the following technical solution: the computer-readable storage medium stores instructions, and when the computer-readable storage medium reads the instructions, it causes the computer-readable storage medium to execute the steps of the risk stratification and decision-making method based on the interaction of a large language model and a risk model as described in the present invention.

[0024] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0025] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 is a flowchart illustrating an embodiment of a risk stratification and decision-making method based on the interaction of a large language model and a risk model according to the present invention; Figure 2 is a structural diagram illustrating an embodiment of a risk stratification and decision-making system based on the interaction of a large language model and a risk model according to the present invention; Figure 3 is a structural diagram illustrating an embodiment of an electronic device according to the present invention. Detailed Implementation

[0026] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0027] Figure 1 illustrates a flowchart of an embodiment of a risk stratification and decision-making method based on the interaction of a large language model and a risk model provided by the present invention. This risk stratification and decision-making method based on the interaction of a large language model and a risk model can be executed by electronic devices such as terminal devices or servers. The terminal device can be any fixed or mobile terminal such as a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, in-vehicle device, or wearable device. The server can be a single server or a server cluster composed of multiple servers. Any electronic device can implement the risk stratification and decision-making method based on the interaction of a large language model and a risk model by having its processor call computer-readable instructions stored in its memory. As shown in Figure 1, the method includes the following steps: S1, acquiring the patient's multimodal clinical data, and preprocessing and extracting features from the multimodal clinical data to obtain a structured feature matrix.

[0028] Here, "patient" refers to an individual receiving disease diagnosis, risk prediction, or health management in a medical setting; for example, a patient A with cardiovascular disease undergoing regular follow-up. Multimodal clinical data refers to a collection of patient health-related information from different sources and types; for example, patient A's age, blood pressure, blood lipid test results, recent electrocardiogram report, and daily heart rate data recorded by wearable devices. The structured feature matrix refers to standardized data, represented numerically, after cleaning and feature extraction, denoted by matrix X. ; This represents the value of the j-th feature, where each row represents a patient and each column represents a feature; for example, a row vector [55,1,145,92,5.2,3.8] in matrix X represents the age, gender, systolic blood pressure, diastolic blood pressure, total cholesterol level, and low-density lipoprotein level of patient A.

[0029] S2. Input the structured feature matrix into the pre-trained risk prediction model and output a multi-dimensional risk spectrum, which includes the patient's risk probability in multiple clinical outcomes.

[0030] Pre-trained risk prediction models refer to machine learning models pre-trained based on historical clinical data to predict the probability of specific clinical events. Specific risk prediction models include Cox proportional hazards models, logistic regression models, and multilayer perceptron models. Multidimensional risk spectra refer to a set of probabilities reflecting the risks a single patient may face across multiple different clinical outcomes, expressed as vectors. ,in This represents the risk probability of the j-th clinical outcome; for example, the vector R = (0.18, 0.12, 0.08) indicates that patient A has a risk probability of 18%, 12%, and 8% for myocardial infarction, heart failure, and stroke within the next year, respectively. A clinical outcome refers to a specific health event or disease state that a patient may experience; for example, myocardial infarction, heart failure, stroke, or cardiac death. The risk probability refers to the estimated likelihood of a specific clinical outcome occurring, calculated by a risk prediction model; for example, the probability of patient A experiencing myocardial infarction is 0.18 (i.e., 18%).

[0031] S3. Based on the risk probability exceeding the preset threshold in the multi-dimensional risk spectrum, retrieve the corresponding evidence-based evidence from the medical knowledge base.

[0032] The preset threshold refers to a manually set critical value used to determine whether the risk probability is high enough to trigger subsequent operations (such as evidence retrieval); for example, setting the risk probability threshold for myocardial infarction to 0.15. The medical knowledge base refers to a database or collection of documents storing a large amount of structured or unstructured medical knowledge; for example, a database containing ESC guidelines, PubMed abstracts, and Cochrane systematic reviews. Evidence-based evidence refers to authoritative medical knowledge or research conclusions from the medical knowledge base that are relevant to a specific clinical problem; for example, a statement from an ESC guideline: "Elevated low-density lipoprotein levels are an important risk factor for cardiovascular events." S4. Input the structured feature matrix, the multidimensional risk spectrum, and the evidence-based evidence into the large language model, and generate an explanatory report through prompt word engineering.

[0033] Here, "large language model" refers to a large-scale artificial intelligence model capable of understanding and generating natural language; for example, ChatGPT, GPT-4, or similar models specifically fine-tuned for medical literature. "Cue word engineering" refers to techniques that guide a large language model to generate specific, high-quality outputs by carefully designing and constructing input text; for example, constructing text containing a patient's risk spectrum and retrieved evidence, instructing the model to generate an explanatory report. "Explanatory report" refers to text content generated by the large language model that explains and clarifies the risk prediction results; for example, a report stating: "Patient A has a high risk of myocardial infarction, primarily driven by elevated low-density lipoprotein levels and poor blood pressure control, consistent with the situation described in the ESC guidelines." S5. Compare the intervention priority for each clinical outcome in the explanatory report with the risk probability of the corresponding clinical outcome in the multi-dimensional risk spectrum; if the intervention priority for the same clinical outcome is inconsistent with the risk level indicated by the risk probability, dynamically adjust the feature weights of the risk prediction model or update the evidence-based retrieval strategy.

[0034] Intervention priority refers to the ranking of the urgency or importance of recommended interventions for different clinical outcomes; for example, the report recommends prioritizing the risk of myocardial infarction (high priority) followed by the risk of stroke (medium priority). Risk level refers to a qualitative description of the probability of risk; for example, a risk probability of 0.18 is qualitatively classified as "medium risk". Feature weight refers to the importance coefficient of each input feature in the risk prediction model for the final prediction result; for example, in a model predicting myocardial infarction, the feature weight of low-density lipoprotein (LDL) is 0.7, and the feature weight of systolic blood pressure is 0.6. Retrieval strategy refers to the methods and rules used when searching for and obtaining relevant evidence-based evidence from medical knowledge bases; for example, using "myocardial infarction AND low-density lipoprotein" as keywords to perform vector similarity retrieval in the FAISS index.

[0035] In S5, specifically: risk probability values ​​are mapped to qualitative risk levels according to preset rules, while the intervention priorities for the same clinical outcome are extracted from the large language model output in the interpretive report. The risk level corresponding to the intervention priority is compared with the risk level mapped from the risk probability. If the comparison results for the same clinical outcome are inconsistent, a dynamic adjustment mechanism is triggered. The dynamic adjustment mechanism selects an adjustment strategy based on the specific content of the inconsistency: if it is determined that there is a deviation between the feature importance assessment of the risk model and clinical knowledge, the weights of features related to the clinical outcome in the risk prediction model are fine-tuned using the backpropagation algorithm; if it is determined that insufficient knowledge retrieval scope or depth leads to insufficient evidence support, the evidence-based evidence retrieval strategy is expanded or modified; for example, adding new keywords extracted from the interpretive report to the search query or adjusting the similarity matching threshold. The above process ensures that the outputs of the risk prediction model and the large language model tend to be consistent, thereby improving the accuracy and credibility of the overall judgment.

[0036] S6. Based on the dynamically adjusted multi-dimensional risk spectrum and preset stratification thresholds, the patient is stratified by risk, and personalized intervention suggestions are generated.

[0037] The dynamically adjusted multidimensional risk spectrum refers to the updated set of risk probabilities after adjustment through an interactive feedback mechanism; for example, after adjusting the model weights based on feedback, the recalculated risk probability of myocardial infarction for patient A changes from 0.18 to 0.16. The preset stratification threshold refers to the set of critical values ​​used to classify patients into different risk levels based on their risk probability values. Indicate; for example, set =0.1, =0.2, the risk probability in the multidimensional risk spectrum R is compared with The comparison categorizes patients into low-risk, high-risk, intermediate-risk, or high-risk groups. Personalized intervention recommendations refer to customized treatment plans generated based on the patient's specific risk stratification and clinical condition; for example, for intermediate-risk patient A, it is recommended to intensify lipid-lowering therapy, strengthen blood pressure monitoring, and implement lifestyle interventions.

[0038] S7. Based on the intervention measures implemented by doctors and patient feedback data, the risk prediction model and the large language model are dynamically optimized to achieve closed-loop iterative updates.

[0039] Intervention measures refer to the actual medical actions taken by doctors for patients based on system recommendations; for example, a doctor adjusts the dosage of statins for patient A and recommends that they increase aerobic exercise. Patient feedback data refers to new health-related data collected from patients after the implementation of intervention measures; for example, patient A's reported symptoms, newly measured blood pressure, and re-examined blood lipid levels one month after the intervention.

[0040] Specifically, in S7: new patient data after intervention is collected and integrated with the original multimodal clinical data and final clinical outcomes to form a new training sample set. For the risk prediction model, an incremental learning algorithm is used to calculate the prediction error loss using the new sample set and update the model parameters and feature weights using gradient descent to optimize its prediction accuracy. For the large language model, explanatory reports are compared with the interventions actually implemented by physicians. Reinforcement learning techniques are used to adjust its prompting strategy and generation logic to improve the consistency between the output and clinical practice. The optimized risk prediction model and large language model are then re-integrated into the risk stratification and decision-making process to process subsequent patient data, thus forming a closed-loop iterative update cycle for prediction, intervention, feedback, and optimization.

[0041] The technical solution in this embodiment solves the problems of insufficient interpretability and disconnection from the knowledge base in traditional risk prediction models through the interaction of a large language model and a risk model. It realizes dynamic risk prediction, generation of interpretive reports and personalized intervention suggestions. At the same time, it optimizes the model based on doctor feedback and patient data in a closed loop, thereby improving the clinical application value and decision-making accuracy.

[0042] In one alternative approach, the steps of preprocessing and feature extraction of the clinical data to obtain a structured feature matrix include: processing missing values ​​in the clinical data using multiple imputation or KNN imputation methods, and removing or correcting outliers in the clinical data based on statistical distribution or medical thresholds to obtain preprocessed clinical data.

[0043] Multiple imputation refers to a method for handling missing data by constructing multiple imputed values ​​to estimate the missing value; for example, using other indicators of patient A (such as age, gender, other test values) to predict and fill in the missing high-density lipoprotein value. KNN imputation refers to a missing value handling method based on the K-nearest neighbor algorithm, filling the missing value with the mean or median of values ​​from similar cases; for example, finding the 10 cases most similar to patient A and filling in A's missing total cholesterol value with the median of their total cholesterol values. Statistical distribution refers to the statistical distribution pattern of data, often used to identify outliers; for example, based on the normal distribution of systolic blood pressure values ​​across all patients, values ​​exceeding three standard deviations from the mean are considered abnormal. Medical thresholds refer to the critical values ​​for determining whether an indicator is normal or not, set based on medical consensus or clinical guidelines; for example, according to guidelines, a low-density lipoprotein level above 4.1 mmol / L is considered abnormal.

[0044] The preprocessed clinical data are subjected to feature screening using principal component analysis or LASSO method to obtain the structured feature matrix.

[0045] Principal component analysis (PCA) refers to a dimensionality reduction method that transforms original features into a set of linearly uncorrelated new features (principal components) through linear transformation; for example, eight related cardiovascular indicators can be transformed into three uncorrelated principal component scores using PCA. The LASSO method refers to a regression method that allows feature selection during model construction; for example, LASSO regression can be used to select the 10 most important features for predicting myocardial infarction from 30 clinical features.

[0046] Among the above-mentioned optional methods, missing values ​​are further processed by multiple imputation or KNN imputation, and features are selected by principal component analysis or LASSO to improve the quality and accuracy of the structured feature matrix and enhance the reliability of the input data of the risk model.

[0047] In one alternative approach, the multidimensional risk spectrum is also visualized using a radar chart or heat map.

[0048] A radar chart is a visual chart that displays risk values ​​from multiple dimensions in a spiderweb-like pattern on a single two-dimensional graph; for example, a pentagonal radar chart can simultaneously display the risk probability of patient A for five cardiovascular events. A heatmap is a matrix-style visual chart that uses color intensity to represent numerical values; for example, a table with a color gradient from green (low risk) to red (high risk) can be used to display the risk of different cardiovascular outcomes in different patients within a patient cohort.

[0049] Among the above-mentioned optional methods, a multi-dimensional risk spectrum can be visually displayed through radar charts or heat maps, enabling clinicians to quickly identify high-risk areas and develop targeted intervention measures.

[0050] In one alternative approach, the step of retrieving corresponding evidence-based evidence from a medical knowledge base includes: retrieving evidence-based evidence corresponding to each clinical outcome for which the risk probability exceeds a preset threshold from the medical knowledge base using a vector retrieval method based on the FAISS index.

[0051] Here, FAISS refers to Facebook AI Similarity Search, a library for efficient similarity searching and dense vector clustering; for example, FAISS can be used to build vector indexes for document summaries in a medical knowledge base to enable fast retrieval.

[0052] Among the above-mentioned optional methods, a vector retrieval method based on the FAISS index is further adopted to quickly and accurately retrieve evidence-based evidence associated with high-risk outcomes in the medical knowledge base, thereby improving knowledge retrieval efficiency and matching accuracy.

[0053] In one alternative approach, the medical knowledge base includes at least one of evidence-based medicine guidelines, meta-analyses, randomized controlled trial literature, and professional databases.

[0054] Evidence-based medicine guidelines refer to authoritative recommendations developed based on the best current scientific evidence to guide clinical practice; for example, the ESC (European Society of Cardiology) guidelines for the prevention of cardiovascular diseases. Meta-analysis refers to a research method that statistically combines and synthesizes the results of multiple independent studies; for example, a meta-analysis article that combined the results of 15 randomized controlled trials to evaluate the effect of aspirin on primary prevention of cardiovascular disease. Randomized controlled trial literature refers to scientific literature that records the process and results of clinical trials designed according to randomization and control principles; for example, an RCT paper published in *JAMA* comparing the effects of two antihypertensive drugs on the prevention of cardiovascular events. Specialized databases refer to databases that contain academic literature and data in specific professional fields; for example, PubMed, a literature database in the biomedical field, or the Cochrane Library, an evidence-based medicine database.

[0055] Among the above-mentioned optional approaches, further integration of various medical knowledge sources such as evidence-based medicine guidelines, meta-analysis, and randomized controlled trial literature can enrich the knowledge base content and enhance the interpretability of prediction results and decision support capabilities.

[0056] In one alternative approach, the step of inputting the structured feature matrix, the multidimensional risk spectrum, and the evidence-based evidence into a large language model and generating an interpretive report through cue word engineering includes: constructing a target instruction through the cue word engineering to cause the large language model to execute the target instruction; analyzing the correlation between the risk probability in the multidimensional risk spectrum and the evidence-based evidence; and generating the interpretive report, which includes attribution of risk drivers, explanation of evidence source, and intervention priority for each clinical outcome.

[0057] The target instruction refers to the text input constructed through prompt word engineering to specifically guide the large language model in performing a specific task; for example, the input text instruction might be: "You are a cardiovascular specialist. Based on the following patient risk data and medical evidence, please generate a risk interpretation report, identifying the main risk drivers and citing evidence." Risk driver attribution refers to the identification and explanation of key clinical features or indicators that lead to increased risk in the interpretive report; for example, the report might state: "The main drivers of this patient's increased risk of myocardial infarction are poorly controlled low-density lipoprotein levels and long-term hypertension." Evidence citation refers to clearly indicating the source of the cited medical advice or conclusions in the interpretive report.

[0058] Among the above-mentioned optional methods, the large language model is further guided by prompt word engineering to comprehensively analyze the multi-dimensional risk spectrum and evidence-based evidence, and generate an interpretive report that includes risk driver attribution, evidence tracing and intervention priority, thereby improving the interpretability of decision support.

[0059] In one alternative approach, the step of risk stratifying the patient based on a dynamically adjusted multidimensional risk spectrum and a preset stratification threshold includes: comparing the risk probability of each clinical outcome in the dynamically adjusted multidimensional risk spectrum with the preset stratification threshold, and classifying the patient into a low-risk group, a medium-risk group, or a high-risk group.

[0060] The low-risk group refers to the patient group with the lowest probability of achieving the target clinical outcome, as defined by risk stratification thresholds; for example, patients with a less than 10% risk of developing myocardial infarction in the next year. The intermediate-risk group refers to the patient group with a moderate probability of achieving the target clinical outcome, as defined by risk stratification thresholds; for example, patients with a 10% to 20% risk of developing myocardial infarction in the next year. The high-risk group refers to the patient group with the highest probability of achieving the target clinical outcome, as defined by risk stratification thresholds; for example, patients with a more than 20% risk of developing myocardial infarction in the next year.

[0061] To better illustrate the technical solution of this embodiment, the following example is used for explanation: S10: Obtain multimodal clinical data of patient A, including age 55 years, gender male, systolic blood pressure 145 mmHg, diastolic blood pressure 92 mmHg, total cholesterol 5.2 mmol / L, low-density lipoprotein 3.8 mmol / L, and recent electrocardiogram report. Missing values ​​in the data are processed using the KNN imputation method, and outliers are removed based on medical thresholds. After feature selection using the LASSO method, a structured feature matrix X=[55,1,145,92] is obtained. [5.2, 3.8]; S20: Input the structured feature matrix X into the pre-trained logistic regression risk prediction model, and output the multi-dimensional risk spectrum R=(0.18, 0.12, 0.08), representing the risk probabilities of myocardial infarction, heart failure, and stroke, respectively; S30: Based on the preset threshold of 0.15, determine that the risk of myocardial infarction requires special attention, and use the FAISS index to retrieve evidence-based evidence K related to low-density lipoprotein from the ESC guidelines in the medical knowledge base; S40: Input the structured feature matrix X, the multi-dimensional risk spectrum R, and the evidence-based evidence K into the large language model, and through... The prompt word engineering generates an explanatory report E, which indicates that the risk driver for myocardial infarction is elevated low-density lipoprotein and the intervention priority is "high." However, the intervention priority corresponding to a stroke risk probability of 0.08 is misclassified as "medium." S50: Comparing the risk levels mapped from intervention priority to risk probability reveals an inconsistency between the stroke risk level of "low risk" and the intervention priority of "medium," triggering a dynamic adjustment mechanism. Due to insufficient retrieval strategy, the keyword "stroke prevention" is added to the search query, and the similarity threshold is adjusted. A re-search yields more relevant evidence. S60: Based on... The updated evidence is used to generate an interpretive report, which corrects the stroke intervention priority to "low". Based on the adjusted risk spectrum and stratification threshold θ=(0.1,0.2), patients are classified into intermediate-risk groups, and personalized intervention recommendations for enhanced lipid-lowering and blood pressure monitoring are generated. S70: When the doctor adjusts the statin dosage of patient A according to the personalized intervention recommendations, and patient feedback data is collected one month later showing that patient A's blood pressure has dropped to the normal range, the risk prediction model and the large language model are dynamically optimized using the intervention measures implemented by the doctor and the patient feedback data, completing the closed-loop iterative update.

[0062] Among the above-mentioned optional methods, the adjusted multidimensional risk spectrum is further compared with the preset stratification threshold to classify patients into low, medium and high-risk groups, providing a clear basis for risk level classification for clinical intervention.

[0063] Figure 2 shows a schematic diagram of an embodiment of a risk stratification and decision-making system 200 based on the interaction of a large language model and a risk model provided by the present invention. As shown in Figure 2, the system 200 includes: a data acquisition module 210, used to acquire multimodal clinical data of patients, and preprocess and extract features from the multimodal clinical data to obtain a structured feature matrix; a risk prediction module 220, used to input the structured feature matrix into a pre-trained risk prediction model and output a multidimensional risk spectrum, the multidimensional risk spectrum including the risk probability of patients on multiple clinical outcomes; a knowledge enhancement module 230, used to retrieve corresponding evidence-based evidence from a medical knowledge base based on the risk probability exceeding a preset threshold in the multidimensional risk spectrum; and a model interaction module 240, used to input the structured feature matrix, the multidimensional risk spectrum, and the evidence-based evidence into a large language model, and generate solutions through prompt word engineering. The system includes: an interpretive report; a dynamic adjustment module 250, used to compare the intervention priority for each clinical outcome in the interpretive report with the risk probability of the corresponding clinical outcome in the multi-dimensional risk spectrum; if the intervention priority for the same clinical outcome is inconsistent with the risk level indicated by the risk probability, the feature weights of the risk prediction model are dynamically adjusted or the evidence retrieval strategy is updated; a risk stratification module 260, used to stratify the patient's risk based on the dynamically adjusted multi-dimensional risk spectrum and a preset stratification threshold, and generate personalized intervention suggestions; and a closed-loop feedback module 270, used to dynamically optimize the risk prediction model and the large language model based on the intervention measures implemented by the doctor and patient feedback data, to achieve closed-loop iterative updates.

[0064] In one optional manner, the data acquisition module 210 is specifically used to: process the missing values ​​in the clinical data using multiple imputation or KNN imputation methods, and remove or correct the outliers in the clinical data based on statistical distribution or medical thresholds to obtain preprocessed clinical data; and perform feature screening on the preprocessed clinical data using principal component analysis or LASSO methods to obtain the structured feature matrix.

[0065] In an alternative approach, the system further includes a visualization module; the visualization module is used to visualize the multidimensional risk spectrum using a radar chart or heat map.

[0066] In one alternative approach, the knowledge enhancement module 230 is specifically used to: retrieve evidence-based evidence corresponding to each clinical outcome whose risk probability exceeds a preset threshold from the medical knowledge base using a vector retrieval method based on the FAISS index.

[0067] In one alternative approach, the medical knowledge base includes at least one of evidence-based medicine guidelines, meta-analyses, randomized controlled trial literature, and professional databases.

[0068] In an alternative approach, the model interaction module 240 is specifically configured to: construct target instructions through the prompt word engineering to enable the large language model to execute the target instructions, analyze the correlation between risk probabilities in the multidimensional risk spectrum and the evidence-based evidence, and generate the explanatory report containing risk driver attribution, evidence source explanation and intervention priority for each clinical outcome.

[0069] In one alternative approach, the risk stratification module 260 is specifically used to: compare the risk probability of each clinical outcome in the dynamically adjusted multidimensional risk spectrum with the preset stratification threshold, and classify the patient into a low-risk group, a medium-risk group, or a high-risk group.

[0070] It should be noted that the beneficial effects of the risk stratification and decision-making system 200 based on the interaction of a large language model and a risk model provided in the above embodiments are the same as those of the risk stratification and decision-making method based on the interaction of a large language model and a risk model, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0071] The risk stratification and decision-making system 200 based on the interaction of a large language model and a risk model of the present invention can be a computer program (including program code) running on a computer device. For example, the risk stratification and decision-making system 200 based on the interaction of a large language model and a risk model of the present invention is an application software that can be used to execute the corresponding steps in the risk stratification and decision-making method based on the interaction of a large language model and a risk model of the present invention.

[0072] In some embodiments, the risk stratification and decision-making system 200 based on the interaction of a large language model and a risk model of the present invention can be implemented in a combination of hardware and software. As an example, the risk stratification and decision-making system 200 based on the interaction of a large language model and a risk model of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the risk stratification and decision-making method based on the interaction of a large language model and a risk model of the present invention. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0073] The modules described in the embodiments of this invention can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.

[0074] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the aforementioned risk stratification and decision-making methods based on the interaction of a large language model and a risk model. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the risk stratification and decision-making method based on the interaction of a large language model and a risk model as shown in any embodiment of the present invention by calling the computer program.

[0075] In one optional embodiment, an electronic device is provided, as shown in FIG3. The electronic device 4000 shown in FIG3 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0076] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0077] Bus 4002 may include a pathway for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent bus 4002 in Figure 3, but this does not mean that there is only one bus or one type of bus.

[0078] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0079] The memory 4003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.

[0080] Among them, electronic devices can also be terminal devices. A terminal device can be any terminal device that can install applications and access web pages through applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0081] It should be noted that the electronic device shown in Figure 3 is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0082] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned risk stratification and decision-making methods based on the interaction of a large language model and a risk model.

[0083] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0084] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned risk stratification and decision-making method based on the interaction of a large language model and a risk model.

[0085] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0086] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0087] The computer-readable storage medium provided in this invention can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0088] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0089] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0090] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0091] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0092] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A risk stratification and decision-making method based on the interaction of a large language model and a risk model, characterized in that, include: Acquire multimodal clinical data of patients, and preprocess and extract features from the multimodal clinical data to obtain a structured feature matrix; The structured feature matrix is ​​input into a pre-trained risk prediction model, which outputs a multi-dimensional risk spectrum, including the patient's risk probability in multiple clinical outcomes. Based on the risk probabilities exceeding a preset threshold in the multidimensional risk spectrum, corresponding evidence-based evidence is retrieved from the medical knowledge base; the structured feature matrix, the multidimensional risk spectrum, and the evidence-based evidence are input into a large language model, and an explanatory report is generated through prompt word engineering; the intervention priority for each clinical outcome in the explanatory report is compared with the risk probabilities of the corresponding clinical outcomes in the multidimensional risk spectrum for consistency. If the intervention priority for the same clinical outcome is inconsistent with the risk level indicated by the risk probability, the feature weights of the risk prediction model shall be dynamically adjusted or the evidence retrieval strategy shall be updated. Based on the dynamically adjusted multi-dimensional risk spectrum and preset stratification thresholds, the patients are stratified by risk, and personalized intervention suggestions are generated. Based on the intervention measures implemented by doctors and patient feedback data, the risk prediction model and the large language model are dynamically optimized to achieve closed-loop iterative updates.

2. The risk stratification and decision-making method based on the interaction of a large language model and a risk model according to claim 1, characterized in that, The steps of preprocessing and feature extraction of the clinical data to obtain a structured feature matrix include: processing missing values ​​in the clinical data using multiple imputation or KNN imputation methods, and removing or correcting outliers in the clinical data based on statistical distribution or medical thresholds to obtain preprocessed clinical data; and performing feature filtering on the preprocessed clinical data using principal component analysis or LASSO methods to obtain the structured feature matrix.

3. The risk stratification and decision-making method based on the interaction of a large language model and a risk model according to claim 1, characterized in that, Also includes: The multi-dimensional risk spectrum can be visualized using radar charts or heat maps.

4. The risk stratification and decision-making method based on the interaction of a large language model and a risk model according to claim 1, characterized in that, The steps for retrieving corresponding evidence-based evidence from the medical knowledge base include: using a vector retrieval method based on the FAISS index from the medical knowledge base to retrieve evidence-based evidence corresponding to each clinical outcome whose risk probability exceeds a preset threshold.

5. The risk stratification and decision-making method based on the interaction of a large language model and a risk model according to claim 4, characterized in that, The medical knowledge base includes at least one of the following: evidence-based medicine guidelines, meta-analysis, randomized controlled trial literature, and professional databases.

6. The risk stratification and decision-making method based on the interaction of a large language model and a risk model according to claim 1, characterized in that, The steps of inputting the structured feature matrix, the multidimensional risk spectrum, and the evidence-based evidence into a large language model and generating an interpretive report through cue word engineering include: constructing a target instruction through cue word engineering to enable the large language model to execute the target instruction; analyzing the correlation between the risk probability in the multidimensional risk spectrum and the evidence-based evidence; and generating the interpretive report, which includes attribution of risk drivers, explanation of evidence source, and intervention priority for each clinical outcome.

7. The risk stratification and decision-making method based on the interaction of a large language model and a risk model according to claim 1, characterized in that, The step of risk stratification of the patient based on the dynamically adjusted multidimensional risk spectrum and the preset stratification threshold includes: comparing the risk probability of each clinical outcome in the dynamically adjusted multidimensional risk spectrum with the preset stratification threshold, and classifying the patient into a low-risk group, a medium-risk group, or a high-risk group.

8. A risk stratification and decision-making system based on the interaction of a large language model and a risk model, characterized in that, include: The data acquisition module is used to acquire patients' multimodal clinical data, and to preprocess and extract features from the multimodal clinical data to obtain a structured feature matrix; The risk prediction module is used to input the structured feature matrix into a pre-trained risk prediction model and output a multi-dimensional risk spectrum, which includes the patient's risk probability in multiple clinical outcomes. The knowledge enhancement module is used to retrieve corresponding evidence-based evidence from the medical knowledge base based on the risk probability exceeding a preset threshold in the multi-dimensional risk spectrum. The model interaction module is used to input the structured feature matrix, the multi-dimensional risk spectrum, and the evidence-based evidence into the large language model, and generate an explanatory report through prompt word engineering; The dynamic adjustment module is used to compare the intervention priority for each clinical outcome in the interpretive report with the risk probability of the corresponding clinical outcome in the multidimensional risk spectrum. If the intervention priority for the same clinical outcome is inconsistent with the risk level indicated by the risk probability, the feature weights of the risk prediction model shall be dynamically adjusted or the evidence retrieval strategy shall be updated. The risk stratification module is used to stratify the patient's risk based on the dynamically adjusted multi-dimensional risk spectrum and preset stratification thresholds, and generate personalized intervention suggestions. The closed-loop feedback module is used to dynamically optimize the risk prediction model and the large language model based on the intervention measures implemented by doctors and patient feedback data, so as to achieve closed-loop iterative updates.

9. An electronic device, characterized in that, The electronic device includes a processor coupled to a memory, the memory storing at least one computer program, which is loaded and executed by the processor to enable the electronic device to implement the risk stratification and decision-making method based on the interaction of a large language model and a risk model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to enable the computer-readable storage medium to implement the risk stratification and decision-making method based on the interaction of a large language model and a risk model as described in any one of claims 1 to 7.

Images