Heart failure auxiliary diagnosis and treatment system, method and medium based on large language model agent

Abstract

The present application relates to the technical field of medical information processing, and discloses a heart failure auxiliary diagnosis and treatment system and method based on a large language model agent, a medium, comprising: a multi-modal data acquisition module for acquiring multi-modal health data of a patient; an HFpEF screening and diagnosis auxiliary agent embedded with a preset diagnosis standard of heart failure with preserved ejection fraction, for automatically performing diagnosis screening of HFpEF and outputting information containing a diagnosis conclusion; an etiological typing and risk assessment agent preset with a clinical feature set for phenotype clustering and prognosis evaluation, for receiving a diagnosis conclusion indicating that the patient is diagnosed as HFpEF, performing phenotype clustering analysis to identify the etiological subtype, and outputting a prognosis risk assessment result; a personalized treatment decision recommendation agent preset with a mapping rule of etiological subtype and drug recommendation scheme and an evidence-based medicine knowledge base, for determining a candidate drug set and generating a personalized drug treatment scheme. A complete auxiliary diagnosis and treatment closed loop from screening to treatment decision is realized.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to a heart failure auxiliary diagnosis and treatment system, method, and medium based on a large language model intelligent agent. Background Technology

[0002] Heart failure with preserved ejection fraction (HFpEF) is a complex clinical syndrome characterized by a left ventricular ejection fraction (LVEF) ≥50% but with typical heart failure symptoms and signs, accounting for nearly 50% of all heart failure patients.

[0003] Currently, the clinical diagnosis and treatment of HFpEF faces the following multiple challenges: First, diagnosis is difficult and the rate of missed diagnosis is high. HFpEF lacks a single specific diagnostic indicator, and its diagnosis relies on a comprehensive assessment of multidimensional information such as clinical manifestations, echocardiographic parameters, and natriuretic peptide levels. The internationally accepted H2FPEF and HFA-PEFF scoring systems have limited sensitivity in Asian populations, and in clinical practice, approximately 90% of patients who meet the diagnostic criteria never receive a formal diagnosis.

[0004] Secondly, the disease exhibits high heterogeneity and treatment decisions are complex. HFpEF is essentially a systemic inflammatory state driven by multiple comorbidities, involving various pathophysiological mechanisms such as obesity-inflammation, diabetes, and chronic kidney disease. The treatment pathways corresponding to different etiological subtypes differ significantly, and traditional heart failure drug treatments have limited efficacy, necessitating precise management strategies based on etiological subtyping.

[0005] Furthermore, long-term patient management is inadequate. HFpEF is a chronic, progressive disease that requires continuous symptom monitoring, medication adjustments, and lifestyle interventions. The existing healthcare system lacks effective outpatient management tools, leading to repeated hospitalizations due to disease progression.

[0006] In recent years, the application of Large Language Models (LLMs) in the medical field has made significant progress. Existing research has explored LLM-based intelligent agent frameworks for hypertension management (such as the Cascade Framework), which, through the combination of Retrieval Enhanced Generation (RAG) and structured reasoning chains, has achieved good results in blood pressure classification, risk factor stratification, and clinical decision support. In the field of heart failure, research has also proposed auxiliary decision generation systems based on multimodal data fusion and knowledge graphs.

[0007] However, existing technologies still have the following shortcomings: First, most existing models are designed for single tasks and lack a comprehensive intelligent agent system that covers the entire chain of screening, diagnosis and classification, treatment decision-making, and long-term management.

[0008] Second, the lack of a mechanism to deeply embed the latest clinical guidelines (such as the 2025 Chinese HFpEF expert consensus) into the agent's decision-making process leads to insufficient transparency and interpretability of decisions.

[0009] Third, existing systems have failed to systematically apply the semantic understanding and reasoning capabilities of large language models to the complex scenario of HFpEF etiological classification and individualized treatment.

[0010] Therefore, there is an urgent need for an intelligent auxiliary system that can systematically solve the above problems and cover the entire diagnosis and treatment process of HFpEF. Summary of the Invention

[0011] The present invention aims to solve the technical problems of high missed diagnosis rate, difficult classification, complex decision-making and weak patient management in the clinical diagnosis and treatment of HFpEF.

[0012] In a first aspect, to address the aforementioned technical problems, a heart failure assisted diagnosis and treatment system based on a large language model intelligent agent is provided, comprising: The multimodal data acquisition module is configured to acquire the patient's multimodal health data; The HFpEF screening and diagnosis assistance agent has a preset ejection fraction retention heart failure diagnostic criteria embedded as inference logic. It is configured to parse the multimodal health data through chain reasoning to automatically perform HFpEF diagnosis screening and output information containing diagnostic conclusions. The etiology subtyping and risk assessment agent is pre-set with a set of clinical features for phenotypic clustering and prognostic assessment. It is configured to receive a diagnostic conclusion indicating that a patient is diagnosed with HFpEF, extract the values ​​of each feature in the set of clinical features from the multimodal health data, perform phenotypic clustering analysis based on the values ​​to identify the etiological subtype, and perform prognostic risk assessment based on the set of clinical features to output the prognostic risk assessment results. The personalized treatment decision-making and recommendation intelligent agent is pre-set with an evidence-based medicine knowledge base containing mapping rules between etiological subtypes and drug recommendation schemes, as well as drug dosage adjustment rules. It is configured to receive the etiological subtype and the prognostic risk assessment results, determine a set of candidate drugs based on the mapping rules, and make applicability judgments and dosage adjustments for each candidate drug according to the patient's personalized parameters to generate a personalized drug treatment plan. The HFpEF screening and diagnosis assistance agent, the etiology classification and risk assessment agent, and the personalized treatment decision recommendation agent are all driven by a large language model and together constitute a multi-agent collaborative workflow with serial data dependencies.

[0013] Furthermore, the multimodal health data includes structured data, unstructured text data, and medical imaging data, wherein: The structured data includes at least demographic information, vital signs, laboratory test results, and information on comorbidities; The unstructured text data includes at least the chief complaint, present illness, physical examination description, past medical history, and echocardiogram report text description from the electronic medical record; The medical imaging data includes at least echocardiogram images.

[0014] Furthermore, the diagnostic criteria include: (a) Presence of symptoms and / or signs of heart failure; (b) Left ventricular ejection fraction (LVEF) ≥ 50%; (c) Presence of left ventricular diastolic dysfunction; And / or, Elevated left ventricular filling pressure consistent with evidence of cardiac structural and / or functional abnormalities; The evidence of cardiac structural and / or functional abnormalities includes at least one of the following: left atrial enlargement, left ventricular hypertrophy, elevated E / e' ratio, pulmonary hypertension estimated by echocardiography, and elevated natriuretic peptide levels.

[0015] Furthermore, the HFpEF screening and diagnostic auxiliary intelligent agent is also configured as follows: Parse unstructured echocardiogram report text to extract features related to diastolic function to help determine evidence of cardiac structural and / or functional abnormalities; Automatically calculate H2FPEF score or HFA-PEFF score and output probability stratification results.

[0016] Furthermore, the etiological subtyping and risk assessment intelligent agent is specifically configured as follows: Phenotypic clustering analysis of patients is performed using pre-trained machine learning clustering methods to identify the etiological subtypes to which they belong; wherein, the clustering method includes K-means clustering or XGBoost-based classification; the etiological subtypes include at least one of obesity-inflammatory type, cardiometabolic type, hypertension-cardiac remodeling type, pulmonary vascular type, and aging-frailty type; The contribution of each feature in the clinical feature set to prognostic risk was calculated and ranked using the SHAP analysis tool, and the risk of readmission to heart failure and / or death in patients at a preset time was assessed. The clinical feature set includes at least one of the following: natriuretic peptide level, estimated glomerular filtration rate, body mass index, history of atrial fibrillation, history of diabetes, history of hypertension, left atrial volume index, E / e' ratio, and pulmonary artery systolic blood pressure.

[0017] Furthermore, the evidence-based medicine knowledge base is constructed using retrieval enhancement generation technology; The drug dosage adjustment rules include: adjusting the diuretic dose or angiotensin receptor neprilysin inhibitor dose when the estimated glomerular filtration rate is lower than a preset threshold, and triggering a potassium supplementation regimen or adjusting the drug combination when serum potassium is lower than a preset threshold.

[0018] Furthermore, it also includes: The decision transparency and reasoning path recording module is configured to record the complete reasoning chain of each agent from input data to output conclusion; the reasoning chain includes the guidelines and evidence sources on which it is based, the data basis and logical judgment of key decision nodes, and interpretability annotations; The long-term management interaction module is configured to provide health management services to patients in the form of a conversational agent.

[0019] Furthermore, the large language model employs retrieval-enhanced generation technology, retrieving relevant information from a preset knowledge base as context before generating output, thereby improving the accuracy and traceability of the output.

[0020] A second aspect of the present invention provides a method for assisted diagnosis and treatment of heart failure based on a large language model intelligent agent, comprising the following steps: Acquire patients' multimodal health data; The embedded preset ejection fraction (HFpEF) diagnostic criteria for heart failure are used as the primary language model to drive the intelligent agent. The agent parses the multimodal health data through chain-like reasoning to automatically perform HFpEF diagnostic screening and outputs information containing diagnostic conclusions. The agent is driven by the second largest language model. It receives the diagnosis conclusion indicating that the patient is diagnosed with HFpEF, and performs phenotypic clustering analysis on the patient's clinical data to identify the etiological subtype and output the prognostic risk assessment results. The intelligent agent is driven by a third language model to receive the etiological subtype and the prognostic risk assessment results, and based on these, combined with a pre-set evidence-based medicine knowledge base, it performs reasoning to generate a personalized drug treatment plan.

[0021] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0022] Compared with the prior art, the present invention has at least the following beneficial effects: This invention systematically solves the problems of single task and lack of full process coverage in existing technologies for HFpEF-assisted diagnosis and treatment by constructing a multi-agent collaborative workflow driven by a large language model.

[0023] The HFpEF screening and diagnosis assistance agent embeds preset diagnostic criteria into chain-like reasoning logic. Each diagnostic condition is decomposed into multiple steps with logical dependencies and the data source is recorded step by step, which improves the transparency and interpretability of the diagnostic process.

[0024] The etiology classification and risk assessment intelligent agent extracts feature values ​​from patient multimodal data based on a preset clinical feature set to perform phenotypic clustering and prognostic assessment, making etiology classification and risk stratification have clear feature basis and reproducibility.

[0025] The personalized treatment decision-making agent uses preset mapping and dosage adjustment rules, combined with the patient's individual parameters, to determine drug suitability and adjust dosage, ensuring that the treatment plan is both in line with evidence-based medicine and suited to the individual characteristics of the patient.

[0026] The three intelligent agents mentioned above form a serial data dependency relationship. The diagnostic conclusion drives the etiology classification, and the classification result drives the treatment decision, realizing a complete auxiliary diagnosis and treatment closed loop from screening to treatment decision, which effectively improves the standardization, accuracy and efficiency of HFpEF diagnosis and treatment. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a system architecture diagram disclosed in this invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] This invention aims to construct a collaborative workflow composed of intelligent agents driven by multiple large language models, which will systematically decompose and intelligently coordinate the entire clinical diagnosis and treatment process of heart failure with preserved ejection fraction (HFpEF)—from screening and diagnosis, etiological classification, risk stratification to personalized treatment decision-making—to achieve complete auxiliary diagnosis and treatment tasks from screening to treatment decision-making.

[0031] In this embodiment, the heart failure auxiliary diagnosis and treatment system based on large language model intelligent agents and ejection fraction preservation mainly includes: a multimodal data acquisition module, an HFpEF screening and diagnosis auxiliary intelligent agent, an etiology classification and risk assessment intelligent agent, a personalized treatment decision recommendation intelligent agent, a long-term patient management interaction module, a decision transparency and reasoning path recording module, an intelligent agent coordination and workflow integration module, and a dynamic update and closed-loop optimization module. A clear serial data dependency relationship is formed between the intelligent agents, that is, diagnostic conclusions drive etiology classification, and classification results drive treatment decisions, thus forming a complete auxiliary diagnosis and treatment closed loop from screening to treatment decision.

[0032] The HFpEF screening and diagnostic aid agent incorporates pre-set diagnostic criteria as its core reasoning logic. Through chain-reasoning technology, it decomposes the complex diagnostic process into a series of logically dependent reasoning steps, recording the data source and reasoning basis for each step. This ensures both diagnostic accuracy and transparency and traceability in decision-making. The etiology subtyping and risk assessment agent has a pre-defined set of clinical features. After extracting feature values ​​from multimodal health data, it performs phenotypic clustering analysis and prognostic risk assessment, ensuring that the subtyping and assessment process has clear feature basis and reproducibility. The personalized treatment decision recommendation agent includes an evidence-based medicine knowledge base containing mapping rules and dosage adjustment rules. It combines patient-specific parameters to assess the applicability of candidate drugs and adjust dosages, ensuring that the generated treatment plan conforms to evidence-based medicine and is tailored to the individual patient characteristics.

[0033] The following combination Figure 1 Each module is described in detail.

[0034] First, the multimodal data acquisition module will be described in detail.

[0035] In this embodiment, the multimodal data acquisition module is responsible for collecting patients' health data from multiple data sources, including structured clinical data, unstructured text data, and medical image data.

[0036] Structured clinical data includes, but is not limited to: demographic information (age, sex, etc.), vital signs (blood pressure, heart rate, body mass index BMI, etc.), laboratory test indicators (NT-proBNP, serum creatinine, serum potassium, estimated glomerular filtration rate eGFR, etc.), and comorbidity information (history of hypertension, history of diabetes, history of atrial fibrillation, history of chronic kidney disease, obesity, etc.).

[0037] Unstructured text data includes, but is not limited to: chief complaints, present illness, physical examination descriptions, past medical history, and text descriptions of echocardiogram reports in electronic medical records.

[0038] Medical imaging data, including but not limited to echocardiogram images.

[0039] After undergoing necessary cleaning, standardization, and vectorization, the aforementioned multimodal health data is transmitted to subsequent intelligent agent modules for further analysis and reasoning.

[0040] Next, a detailed explanation of the HFpEF screening and diagnostic auxiliary agent will be provided.

[0041] In this embodiment, the HFpEF screening and diagnosis auxiliary intelligent agent is the first core intelligent agent in the system workflow, driven by the first large language model. This intelligent agent embeds a preset ejection fraction-preserving heart failure diagnostic standard as inference logic, and is configured to parse multimodal health data to automatically perform HFpEF diagnostic screening and output information containing diagnostic conclusions.

[0042] The first major language model can be a general-purpose base language model, such as DeepSeek-V3, ChatGPT-4o, or other models with comparable capabilities.

[0043] In a preferred embodiment, the preset diagnostic criteria for heart failure with preserved ejection fraction are the diagnostic criteria in the 2025 edition of the "Chinese Expert Consensus on the Diagnosis and Treatment of Heart Failure with Preserved Ejection Fraction".

[0044] The diagnostic criteria include the following core conditions: First, there are symptoms and / or signs of heart failure.

[0045] Second, left ventricular ejection fraction (LVEF) ≥ 50%.

[0046] Third, there is evidence of cardiac structural and / or functional abnormalities consistent with left ventricular diastolic dysfunction and / or elevated left ventricular filling pressure.

[0047] To transform the aforementioned diagnostic criteria from medical text in natural language form into reasoning logic executable by a large language model, this embodiment employs a structured cue word engineering approach. Specifically, for each diagnostic condition, at least two experts with extensive clinical experience in the cardiovascular field independently provide the following: the medical definition of the condition, judgment rules (including numerical thresholds and qualitative criteria), various medical descriptions that may appear in electronic medical records, and evidence priorities and mutually exclusive conditions. Based on the aforementioned expert knowledge, a structured cue word template is constructed.

[0048] Taking the presence of evidence of cardiac structural and / or functional abnormalities in the diagnostic criteria as an example, this condition is further refined in the prompt template into the following sub-conditions that can be judged item by item: left atrial enlargement (left atrial volume index greater than preset value), left ventricular hypertrophy (left ventricular mass index greater than preset value), elevated ratio of early diastolic mitral valve velocity to early diastolic mitral valve annulus velocity (E / e' average greater than 14), pulmonary hypertension estimated by echocardiography, and elevated natriuretic peptide levels (NT-proBNP greater than preset age-adjusted threshold). The prompt template provides corresponding lists of medical descriptions and evidence priorities for each of these sub-conditions.

[0049] For example, for the sub-condition of elevated E / e', the medical descriptions configured in the prompt template can include: expressions from the ultrasound report such as "elevated E / e' ratio," "E / e' > 14," "diastolic dysfunction (grade II or above)," and "increased ratio of early diastolic flow velocity to mitral annulus velocity." For each medical description, its evidence priority in the reasoning process is further configured—direct numerical evidence from quantitative echocardiographic reports has the highest priority, qualitative descriptions from ultrasound descriptive text have the second highest priority, and descriptions from the physician's subjective judgment have the lowest priority.

[0050] After completing the rule-based embedding of diagnostic criteria, the agent uses chain reasoning technology to decompose the complex diagnostic judgment process into a series of logically dependent reasoning steps.

[0051] As those skilled in the art will know, chain reasoning is a technique that guides a large language model to perform logical reasoning step by step. By decomposing a macro task into multiple interrelated sub-tasks, the model can focus on a single judgment target at each step, while recording the reasoning basis for that step.

[0052] In some specific implementations, the chained reasoning for HFpEF diagnosis is constructed as the following sequence of logical steps: The first step is to assess the symptoms and signs of heart failure. Information relevant to the clinical manifestations of heart failure is extracted from multimodal health data, including symptom descriptions in the patient's chief complaint (e.g., dyspnea, decreased exercise tolerance, paroxysmal nocturnal dyspnea, orthopnea, etc.), physical examination findings (e.g., lung rales, jugular venous distension, lower extremity edema, etc.), and NYHA functional classification information. Based on this information, a comprehensive assessment is made to determine whether the patient has heart failure symptoms and / or signs, and the original symptom descriptions and source fields are recorded.

[0053] The second step is LVEF value extraction and judgment. Left ventricular ejection fraction (LVEF) values ​​are extracted from multimodal health data, primarily from quantitative values ​​in echocardiogram reports; if no quantitative values ​​are available, they are inferred from the report text description. After extraction, the LVEF values ​​are compared with a 50% threshold, and the comparison results are output, along with the source of the LVEF values.

[0054] The third step involves the item-by-item evaluation of evidence regarding cardiac structural / functional abnormalities. Following the sub-conditions defined in the prompt template, relevant evidence is retrieved and extracted from multimodal health data item by item. For each sub-condition, the highest priority source of evidence is first identified; if it exists, it is evaluated directly. If the highest priority evidence is missing, it is evaluated from lower priority sources. The evaluation result for each sub-condition is recorded, including the data segment used, the source field, and the evaluation logic.

[0055] The fourth step is to generate a comprehensive diagnostic conclusion. The results of the first three steps are summarized, and a final diagnostic conclusion is generated based on the logical relationship of the diagnostic criteria (all three conditions must be met for a definitive diagnosis). The diagnostic conclusion includes one of the following types: confirmed HFpEF, suspected HFpEF (meets some conditions but the evidence is insufficient), or exclusion of HFpEF (does not meet the necessary conditions).

[0056] In the chain-like reasoning process described above, each step generates an intermediate reasoning record. This intermediate reasoning record includes the step number, judgment content, input data fragment, judgment logic, judgment result, and the data source on which it is based. These intermediate reasoning records together constitute the complete reasoning chain of the agent, providing support for subsequent clinical review and decision traceability.

[0057] In a preferred embodiment, the HFpEF screening and diagnostic aid agent is also configured to automatically calculate the H2FPEF score or HFA-PEFF score and output the probability stratification results.

[0058] The H2FPEF score integrates indicators such as body mass index, age, history of hypertension, history of atrial fibrillation, pulmonary artery systolic blood pressure, and E / e' ratio. Each indicator is assigned a certain score, and the higher the score, the greater the likelihood of HFpEF. The HFA-PEFF score integrates echocardiographic functional indicators, echocardiographic morphological indicators, and biomarker indicators, classifying patients into three levels: low probability, moderate probability, and high probability.

[0059] Furthermore, the HFpEF screening and diagnostic aid is configured to parse unstructured echocardiogram report text, extracting features related to diastolic function to aid in identifying evidence of cardiac structural and / or functional abnormalities. For example, it can extract parameters such as left atrial anteroposterior diameter, left atrial volume index, interventricular septal thickness, left ventricular posterior wall thickness, early diastolic velocity e' of the mitral annulus, E / e' ratio, and peak tricuspid regurgitation velocity from the ultrasound report. This parsing process can be performed using the natural language understanding capabilities of the large language model itself, or it can be combined with named entity recognition technology to improve extraction accuracy.

[0060] Secondly, a detailed explanation of the etiological classification and risk assessment intelligent agent is provided.

[0061] In this embodiment, the etiology subtyping and risk assessment agent is the second core agent in the system workflow, driven by the second largest language model. This agent has a pre-defined set of clinical features for phenotypic clustering and prognostic assessment. Upon receiving a diagnosis indicating that the patient has HFpEF, the agent extracts the values ​​of each feature from the clinical feature set from multimodal health data, performs phenotypic clustering analysis based on these values ​​to identify the patient's etiological subtype, and performs prognostic risk assessment based on the clinical feature set to output the prognostic risk assessment result.

[0062] The etiology classification and risk assessment agent is pre-set with a set of clinical features for phenotypic clustering and prognostic assessment. The selection of this set of clinical features is based on variables that have a significant impact on patient prognosis in a large number of HFpEF clinical studies. It was determined after discussion and statistical screening by clinical experts, and the set of clinical features fully considers the multidimensional pathophysiological characteristics of HFpEF.

[0063] In one specific implementation, the clinical feature set includes at least one of the following: natriuretic peptide level (NT-proBNP or BNP), estimated glomerular filtration rate (eGFR), body mass index (BMI), history of atrial fibrillation, history of diabetes, history of hypertension, left atrial volume index, the ratio of early diastolic flow velocity at the mitral valve orifice to early diastolic flow velocity at the mitral valve annulus (E / e'), and pulmonary artery systolic pressure.

[0064] Those skilled in the art will recognize that natriuretic peptide levels reflect cardiac filling pressure and ventricular wall stress, and are core biomarkers for assessing the severity of heart failure; eGFR reflects renal function status, and renal insufficiency is one of the most common comorbidities in patients with heart failure-preventive embolism (HFpEF), and is also a key factor affecting prognosis; BMI reflects the degree of obesity, and obesity-inflammation is one of the important pathogenesis mechanisms of HFpEF; a history of atrial fibrillation and hypertension are the most common comorbidities of HFpEF, directly associated with specific etiological subtypes; left atrial volume index and E / e' ratio are core ultrasound indicators for assessing left ventricular diastolic function and left atrial pressure; and pulmonary artery systolic pressure reflects pulmonary circulation status and is closely related to pulmonary vascular HFpEF.

[0065] Upon receiving a diagnosis indicating that the patient has HFpEF, the etiology subtyping and risk assessment agent first extracts the values ​​of each feature from the clinical feature set of multimodal health data. Specifically, for structured data, values ​​can be directly extracted from the corresponding fields. For relevant information in unstructured text data, the agent utilizes the natural language understanding capabilities of the second-largest language model for semantic parsing and numerical extraction.

[0066] After feature extraction is completed, the etiology typing and risk assessment agent performs phenotypic clustering analysis based on the extracted feature values. In one specific implementation, this clustering analysis is performed using a pre-trained machine learning clustering method based on the K-means clustering algorithm or the XGBoost classification algorithm.

[0067] Taking K-means clustering as an example, its specific clustering methods include: First, multidimensional clinical data of a large-scale HFpEF patient cohort were collected, and the values ​​of each feature in the above clinical feature set were extracted and standardized as input for clustering.

[0068] Then, the optimal number of clusters is determined using the elbow rule.

[0069] Next, the K-means algorithm is used to cluster the patient samples to maximize the similarity of patient features within the same cluster and the feature differences between different clusters.

[0070] Finally, by analyzing the characteristic distribution patterns of each cluster center, clinical experts assigned corresponding etiological subtype labels to each cluster.

[0071] After pre-training and labeling, the etiology typing and risk assessment agent can identify at least one of the following etiology subtypes: Obesity-inflammatory type – characterized by high BMI, often accompanied by significant metabolic abnormalities and elevated systemic inflammatory markers; Cardiometabolic type – characterized by diabetes and metabolic syndrome as the main driving factors, combined with multiple cardiometabolic risk factors; Hypertension-cardiac remodeling type – characterized by left ventricular concentric remodeling and diastolic dysfunction caused by long-term hypertension; Pulmonary vascular type – characterized by pulmonary hypertension and right ventricular dysfunction; Aging-frailty type – characterized by advanced age, multi-system functional decline, and frailty.

[0072] In the actual reasoning process, the extracted patient feature values ​​are input into the etiology subtyping and risk assessment agent to obtain the etiology subtype classification result to which the patient belongs. At the same time, the second language model provides semantic interpretation of the clustering results—for example, analyzing the distribution of the patient's feature values ​​across various clustering dimensions to explain the main driving features that led to their classification into a specific subtype—making the classification conclusion clinically understandable and acceptable.

[0073] Regarding prognostic risk assessment, the etiological subtyping and risk assessment agent, based on the aforementioned clinical feature set, employs interpretability analysis tools to evaluate the impact of each feature on patient prognosis. In one specific implementation, the agent uses the SHAP analysis tool to calculate and rank the contribution of each feature in the clinical feature set to prognostic risk. As those skilled in the art will recognize, SHAP is a game theory-based interpretability analysis method that quantifies the importance of each feature by calculating its marginal contribution to the model's prediction results.

[0074] Specifically, the etiology classification and risk assessment agent is pre-configured with a risk prediction model trained on a historical HFpEF patient cohort. This risk prediction model takes each feature of the clinical feature set as input and uses heart failure readmission events and / or deaths within one year as prediction endpoints. The SHAP analysis tool calculates the contribution of each feature in the risk prediction model to the prediction outcome for a specific patient; positive values ​​indicate that the feature increases the risk, and negative values ​​indicate that the feature decreases the risk. The contribution values ​​of each feature are sorted from largest to smallest in absolute value, forming a ranking list of key prognostic features.

[0075] Based on the output of the risk prediction model, the agent generates a prognostic risk assessment result. This assessment result includes a stratification of the patient's one-year risk of readmission and death from heart failure, categorized as low, intermediate, or high risk. In one specific implementation, a predicted one-year risk of readmission or death from heart failure is defined as low risk if less than 10%, intermediate risk if between 10% and 30%, and high risk if greater than 30%. This risk stratification result, along with the importance ranking of SHAP features, is used as the output of the agent and passed to the downstream personalized treatment decision-making recommendation agent.

[0076] Furthermore, a detailed explanation of the personalized treatment decision-making and recommendation agent is provided.

[0077] In this embodiment, the personalized treatment decision-making recommendation agent is the third core agent in the system workflow, driven by a third major language model. This agent is pre-loaded with an evidence-based medicine knowledge base containing mapping rules between etiological subtypes and recommended drug regimens, as well as rules for adjusting drug dosage. Upon receiving the etiological subtype and prognostic risk assessment results, the agent determines a set of candidate drugs based on the mapping rules and performs suitability assessments and dosage adjustments for each candidate drug according to the patient's personalized parameters, generating a personalized drug treatment plan.

[0078] The personalized treatment decision-making agent is built upon a pre-defined evidence-based medicine knowledge base, which forms the foundation for its accurate recommendations. This knowledge base is constructed using retrieval-enhanced generation technology, systematically integrating authoritative clinical guidelines, drug databases, and high-quality historical cases. Knowledge sources within the knowledge base include, but are not limited to: the 2025 edition of the "Chinese Expert Consensus on the Diagnosis and Treatment of Heart Failure with Preserved Ejection Fraction," the ESC (European Society of Cardiology) Heart Failure Guidelines, the ACCF / AHA (American College of Cardiology Foundation / American Heart Association) Heart Failure Guidelines, the DrugBank drug database, and high-quality local historical cases reviewed by clinical experts.

[0079] In one specific implementation, the evidence-based medicine knowledge base contains two core rules: the first is the mapping rule between etiological subtypes and recommended drug regimens, and the second is the drug dosage adjustment rule.

[0080] The mapping rule between etiological subtypes and recommended drug regimens establishes a correspondence between etiological subtypes and recommended drug sets. For example, for obesity-inflammatory HFpEF, the mapping rule prioritizes GLP-1 receptor agonists or GIP / GLP-1 dual receptor agonists because these drugs have been shown to improve clinical outcomes in HFpEF patients while achieving weight loss; for hypertension-cardiac remodeling HFpEF, the mapping rule prioritizes optimized antihypertensive drug regimens and SGLT2 inhibitors to address both blood pressure control and heart failure treatment; for cardiometabolic HFpEF, the mapping rule recommends SGLT2 inhibitors and fenestrone to cover the dual needs of metabolic disorders and cardiorenal protection.

[0081] Drug dosage adjustment rules include, but are not limited to: (1) Adjust the dose of drugs excreted by the kidneys according to the eGFR level—for example, when the eGFR is below 30 mL per minute, reduce the recommended dose of angiotensin receptor neprilysin inhibitor (ARNI) to 50 percent of the standard dose.

[0082] (2) When eGFR is below 30 ml / min, adjust the type or dose of diuretic to avoid further damage to kidney function.

[0083] (3) Adjust potassium-sparing diuretics and potassium supplementation regimens according to serum potassium levels. For example, when serum potassium is below 3.5 mmol / L, trigger a potassium supplementation regimen or adjust the drug combination to avoid the risk of hypokalemia.

[0084] (4) When serum potassium is higher than 5.5 mmol / L, it is recommended to reduce or discontinue the use of potassium-sparing diuretics.

[0085] In this embodiment, after receiving the etiological subtype and prognostic risk assessment results, the personalized treatment decision recommendation agent performs the following steps: The first step is to determine the candidate drug set. Based on the received etiological subtypes, the corresponding mapping rules are queried in the evidence-based medicine knowledge base to determine the candidate drug set corresponding to that subtype.

[0086] The second step is to obtain personalized parameters. Personalized parameters of patients are extracted from multimodal health data. These parameters include at least: fluid retention status (based on a comprehensive assessment of physical signs such as weight changes, degree of edema, and lung rales), renal function status (based on indicators such as eGFR and serum creatinine), electrolyte levels (serum potassium, serum sodium, etc.), blood pressure level, liver function status, LVEF value, and BMI value.

[0087] The third step is drug suitability assessment. For each drug in the candidate drug set, a suitability assessment is performed individually. This assessment process includes: checking for contraindications (e.g., limited use of certain renally excreted drugs in cases of severe renal insufficiency, limited use of potassium-sparing diuretics in cases of hyperkalemia, and caution required when using antihypertensive drugs in cases of hypotension); checking for adverse interactions between the drug and the patient's comorbidities and concomitant medications; and assessing the expected benefit-risk ratio of the drug for the patient's specific subtype and risk stratification.

[0088] The fourth step is drug dosage adjustment. For drugs that have passed the suitability assessment, the dosage is calculated according to the pre-set drug dosage adjustment rules and the patient's individual parameters. For example, taking sacubitril / valsartan as an example, if the patient's LVEF ≤ 57% and eGFR ≥ 30 ml / min, then this drug is recommended. The starting dose is determined according to the patient's blood pressure level—those with normal or high blood pressure start with half of the standard dose, and those with low blood pressure start with one-quarter of the standard dose; if eGFR < 30 ml / min, then the recommended dose is reduced or it is recommended to switch to an alternative drug.

[0089] The fifth step is to generate a comprehensive treatment plan. All medications that have passed suitability assessment and dosage adjustments are compiled, prioritized, and a complete personalized treatment plan is formed. This plan includes the medication name, starting dose, dosing frequency, and follow-up monitoring recommendations.

[0090] In a preferred embodiment, the personalized treatment decision recommendation agent also uses retrieval-enhanced generation technology to retrieve clinical guideline recommendations, treatment experiences of historical high-quality cases, or the latest clinical research evidence that are similar to the current patient's condition from the knowledge base in real time when generating treatment plans, as a reference and evidence-based support for the recommended plans.

[0091] In addition, a detailed explanation of the patient long-term management interaction module is provided.

[0092] In this embodiment, the patient long-term management interaction module is driven by a fourth language model, providing health management services to patients in the form of a conversational intelligent agent, including symptom monitoring, health guidance, and early warning services. This extends in-hospital treatment decisions to long-term out-of-hospital management, effectively compensating for the shortcomings of the existing medical system in patient out-of-hospital management.

[0093] In some specific implementations, the patient long-term management interaction module includes at least one of the following functional sub-modules: (a) The symptom periodic collection submodule is used to periodically collect information on changes in patients' symptoms. This submodule proactively initiates symptom inquiries with patients through a conversational interface at preset time intervals. The inquiries include, but are not limited to, the degree of dyspnea, edema, weight changes, and exercise tolerance. Patients can respond using natural language or by selecting options. This submodule records the collected symptom information in a structured manner, forming time-series data on changes in patients' out-of-hospital symptoms.

[0094] (b) The Lifestyle Intervention Recommendation Push Submodule is used to send periodic health education and lifestyle intervention recommendations based on the patient's individual characteristics. This submodule selects matching intervention recommendations from a pre-built health education content library and pushes them to the patient based on the patient's etiological subtype and personalized parameters. The intervention recommendations include, but are not limited to: low-salt diet guidance, fluid management recommendations, and exercise rehabilitation guidance.

[0095] (c) The medication reminder and follow-up appointment reminder submodule is used to remind patients to take their medication, attend follow-up appointments, and complete relevant examinations on time. This submodule sends medication reminders to patients at preset times based on the dosing frequency and timing specified in the medication plan output by the personalized treatment decision-making agent. Simultaneously, according to the follow-up appointment schedule set by the doctor, it sends follow-up appointment reminders to patients as the appointment date approaches, including the appointment time and required examinations.

[0096] (d) The abnormal vital signs early warning submodule is used to perform preliminary analysis of abnormal vital signs reported by patients, generate early warning signals, and recommend medical attention. This submodule has a set of preset threshold standards for abnormal vital signs, such as: rapid weight gain of more than two kilograms in a short period of time, severe dyspnea, new chest pain or palpitations, etc. When the symptom data reported by the patient reaches or exceeds the above thresholds, this submodule automatically generates an early warning signal, notifies the patient and their attending physician through system messages or SMS, and recommends that the patient seek medical attention promptly or undergo emergency treatment according to the severity of the warning.

[0097] The interactions of all the above sub-modules are conducted in the form of natural language dialogue. The fourth language model is configured with appropriate prompt words so that it can communicate with patients in an empathetic and easy-to-understand language style, thereby improving patient participation and compliance.

[0098] Finally, the decision transparency and reasoning path recording module, the agent coordination and workflow integration module, and the dynamic update and closed-loop optimization module are described in detail.

[0099] The decision transparency and reasoning path recording module is responsible for recording the complete reasoning chain of each agent from input data to output conclusion, including the guidelines and evidence sources on which it is based, the data basis and logical judgment of key decision nodes, and interpretable annotations.

[0100] Specifically: For the HFpEF screening and diagnostic assistance agent, this module records the reasoning chain, including: the judgment process for each diagnostic condition—the data segments involved, the judgment logic, and the judgment result; the logical dependencies between each diagnostic condition; and the derivation process of the final diagnostic conclusion. For example, for the judgment of the sub-condition "elevated E / e' ratio" in the third core condition, this module records: the specific E / e' value extracted from the ultrasound report, the comparison logic with the preset threshold, and the judgment result. This complete recording of information allows clinicians to trace the data source and reasoning logic of each diagnostic judgment, enhancing the credibility of the system output.

[0101] For the etiology classification and risk assessment agent, the reasoning chain recorded in this module includes: the specific values ​​and sources of each feature in the clinical feature set extracted from multimodal health data; the feature input and classification output of the clustering model; the contribution ranking results of each feature in the SHAP analysis; and the calculation basis of the risk stratification results.

[0102] For the personalized treatment decision-making agent, the reasoning chain recorded by this module includes: etiological subtype and risk stratification information that triggers treatment recommendations; mapping rules and dosage adjustment rules retrieved from the evidence-based medicine knowledge base; applicability judgment results and dosage calculation process for each candidate drug; and the priority ranking basis for the recommended schemes.

[0103] The complete inference chain record and interpretable annotations described above transform the entire assisted diagnosis process from a black-box model output into a logically clear, transparent, and traceable reasoning process. This provides a solid data foundation for medical quality control, physician review and confirmation, and continuous system optimization and iteration in clinical applications.

[0104] In this embodiment, the intelligent agent coordination and workflow integration module is responsible for scheduling the aforementioned intelligent agents to achieve collaborative work throughout the entire process from screening, diagnosis and classification, treatment recommendation to long-term management, and integrates with the hospital information system or electronic medical record system through API interface.

[0105] In this embodiment, the dynamic update and closed-loop optimization module is used to dynamically update the patient profile and treatment plan based on patient follow-up data, treatment response, and newly generated clinical evidence, forming a closed-loop optimization mechanism for diagnosis and treatment.

[0106] It should be noted that the first, second, third, and fourth language models mentioned above can be different instances of the same basic model, or they can be different models fine-tuned for their respective tasks. The driving model for each agent can be flexibly selected according to the actual deployment scenario. For example, for the HFpEF screening and diagnosis assistance agent, a model fine-tuned on cardiovascular clinical text can be used to improve the accuracy of professional reasoning, while for the patient long-term management interaction module, a model focusing on natural dialogue capabilities can be used to improve the patient interaction experience.

[0107] The system of this invention adopts a modular and scalable architecture design, which can be flexibly deployed in medical institutions of different sizes and types.

[0108] Regarding hardware configuration, the system can be deployed either locally or in the cloud, depending on the hospital's specific needs. For local deployment, the system can be deployed on a server cluster within the hospital, with server configuration determined based on the hospital's average daily outpatient and inpatient volume. Model inference for each agent can run on GPU servers to improve response speed, and data storage can utilize the hospital's existing database and file storage systems.

[0109] In terms of software integration, the intelligent agent coordination and workflow integration module of this system interfaces with the hospital's existing information systems through standardized API interfaces. These information systems include, but are not limited to: Hospital Information System (HIS), Electronic Medical Record System (EMR), Laboratory Information System (LIS), and Medical Image Archiving and Communication System (PACS). The API interfaces adopt medical information standard protocols such as RESTful or HL7 FHIR to ensure the security and interoperability of data transmission.

[0110] In terms of workflow integration, the system supports both automatic and manual triggering modes. The automatic triggering mode initiates the assessment process automatically at preset time points (such as new patient admission, test result reporting, and fixed daily times). The manual triggering mode allows clinicians to initiate assessments at any time on the doctor's workstation interface as needed. After the assessment is completed, the system pushes the results to the doctor's workstation for review and confirmation. Doctors can confirm, modify, or reject the system's output diagnoses, classification results, or treatment plans; all modifications are recorded for continuous system optimization.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A heart failure auxiliary diagnosis and treatment system based on a large language model intelligent agent, characterized in that, include: The multimodal data acquisition module is configured to acquire the patient's multimodal health data; The HFpEF screening and diagnosis assistance agent has a preset ejection fraction retention heart failure diagnostic criteria embedded as inference logic. It is configured to parse the multimodal health data through chain reasoning to automatically perform HFpEF diagnosis screening and output information containing diagnostic conclusions. The etiology subtyping and risk assessment agent is pre-set with a set of clinical features for phenotypic clustering and prognostic assessment. It is configured to receive a diagnostic conclusion indicating that a patient is diagnosed with HFpEF, extract the values ​​of each feature in the set of clinical features from the multimodal health data, perform phenotypic clustering analysis based on the values ​​to identify the etiological subtype, and perform prognostic risk assessment based on the set of clinical features to output the prognostic risk assessment results. The personalized treatment decision-making and recommendation intelligent agent is pre-set with an evidence-based medicine knowledge base containing mapping rules between etiological subtypes and drug recommendation schemes, as well as drug dosage adjustment rules. It is configured to receive the etiological subtype and the prognostic risk assessment results, determine a set of candidate drugs based on the mapping rules, and make applicability judgments and dosage adjustments for each candidate drug according to the patient's personalized parameters to generate a personalized drug treatment plan. The HFpEF screening and diagnosis assistance agent, the etiology classification and risk assessment agent, and the personalized treatment decision recommendation agent are all driven by a large language model and together constitute a multi-agent workflow with serial data dependencies.

2. The heart failure auxiliary diagnosis and treatment system based on a large language model intelligent agent according to claim 1, characterized in that, The multimodal health data includes structured data, unstructured text data, and medical imaging data, wherein: The structured data includes at least demographic information, vital signs, laboratory test results, and information on comorbidities; The unstructured text data includes at least the chief complaint, present illness, physical examination description, past medical history, and echocardiogram report text description from the electronic medical record; The medical imaging data includes at least echocardiogram images.

3. The heart failure auxiliary diagnosis and treatment system based on a large language model intelligent agent according to claim 1, characterized in that, The diagnostic criteria include: (a) Presence of symptoms and / or signs of heart failure; (b) Left ventricular ejection fraction (LVEF) ≥ 50%; (c) Presence of left ventricular diastolic dysfunction; And / or, Elevated left ventricular filling pressure consistent with evidence of cardiac structural and / or functional abnormalities; The evidence of cardiac structural and / or functional abnormalities includes at least one of the following: left atrial enlargement, left ventricular hypertrophy, elevated E / e' ratio, pulmonary hypertension estimated by echocardiography, and elevated natriuretic peptide levels.

4. The heart failure auxiliary diagnosis and treatment system based on a large language model intelligent agent according to claim 3, characterized in that, The HFpEF screening and diagnostic auxiliary intelligent agent is also configured as follows: Parse unstructured echocardiogram report text to extract features related to diastolic function to help determine evidence of cardiac structural and / or functional abnormalities; Automatically calculate H2FPEF score or HFA-PEFF score and output probability stratification results.

5. The heart failure auxiliary diagnosis and treatment system based on a large language model intelligent agent according to claim 1 or 3, characterized in that, The etiological classification and risk assessment intelligent agent is specifically configured as follows: Phenotypic clustering analysis of patients is performed using pre-trained machine learning clustering methods to identify the etiological subtypes to which they belong; wherein, the clustering method includes K-means clustering or XGBoost-based classification; the etiological subtypes include at least one of obesity-inflammatory type, cardiometabolic type, hypertension-cardiac remodeling type, pulmonary vascular type, and aging-frailty type; The contribution of each feature in the clinical feature set to prognostic risk was calculated and ranked using the SHAP analysis tool, and the risk of readmission to heart failure and / or death in patients at a preset time was assessed. The clinical feature set includes at least one of the following: natriuretic peptide level, estimated glomerular filtration rate, body mass index, history of atrial fibrillation, history of diabetes, history of hypertension, left atrial volume index, E / e' ratio, and pulmonary artery systolic blood pressure.

6. The heart failure auxiliary diagnosis and treatment system based on a large language model intelligent agent according to claim 5, characterized in that, The evidence-based medicine knowledge base is constructed using retrieval enhancement generation technology; The drug dosage adjustment rules include: adjusting the diuretic dose or angiotensin receptor neprilysin inhibitor dose when the estimated glomerular filtration rate is lower than a preset threshold, and triggering a potassium supplementation regimen or adjusting the drug combination when serum potassium is lower than a preset threshold.

7. The heart failure auxiliary diagnosis and treatment system based on a large language model intelligent agent according to claim 1, characterized in that, Also includes: The decision transparency and reasoning path recording module is configured to record the complete reasoning chain of each agent from input data to output conclusion; the reasoning chain includes the guidelines and evidence sources on which it is based, the data basis and logical judgment of key decision nodes, and interpretability annotations; The long-term management interaction module is configured to provide health management services to patients in the form of a conversational agent.

8. The heart failure auxiliary diagnosis and treatment system based on a large language model intelligent agent according to claim 1, characterized in that, The large language model employs retrieval-enhanced generation technology, retrieving relevant information from a pre-set knowledge base as context before generating output.

9. A method for assisted diagnosis and treatment of heart failure based on a large language model intelligent agent, characterized in that, Includes the following steps: Acquire patients' multimodal health data; The embedded preset ejection fraction (HFpEF) diagnostic criteria for heart failure are used as the primary language model to drive the intelligent agent. The agent parses the multimodal health data through chain-like reasoning to automatically perform HFpEF diagnostic screening and outputs information containing diagnostic conclusions. The agent is driven by the second largest language model. It receives the diagnosis conclusion indicating that the patient is diagnosed with HFpEF, and performs phenotypic clustering analysis on the patient's clinical data to identify the etiological subtype and output the prognostic risk assessment results. The intelligent agent is driven by a third language model to receive the etiological subtype and the prognostic risk assessment results, and based on these, combined with a pre-set evidence-based medicine knowledge base, it performs reasoning to generate a personalized drug treatment plan.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in claim 9.

Images