A method for diagnosing and predicting acute aortic dissection in a multi-center environment

Abstract

This application provides a method for diagnosing and predicting acute aortic dissection in a multicenter environment, comprising: an agent system acquiring first-level clinical data of a patient and converting it into narrative clinical description information, wherein the clinical description information ends with a diagnostic question, requiring the selection of a suspected or non-target disease; a large language model performing semantic understanding on the clinical description information and generating a first-level risk assessment, selecting a suspected or non-target disease; if the first-level risk assessment indicates a suspected target disease, the agent system terminates subsequent data input and outputs a target disease risk warning; otherwise, the agent system acquires second-level clinical data of the patient and incorporates it into the clinical description information; the large language model generating a third-level risk assessment on the final clinical description information, selecting a suspected or non-target disease, and outputting the final prediction result.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for the diagnosis and prediction of acute aortic dissection in a multicenter setting. Background Technology

[0002] Acute aortic dissection is one of the most catastrophic cardiovascular emergencies in emergency patients. It has a rapid onset, rapid progression, and high mortality rate, and is known as a "time-dependent killer." Risk prediction and early identification of these patients is a highly challenging research topic in the field of emergency medicine. At the same time, in the emergency environment, the large patient flow and the complexity and diversity of their conditions require medical staff to make triage decisions in a very short time, which puts great pressure on the risk assessment process.

[0003] Existing methods typically rely on fixed procedures and pre-defined rules for assessment, often lacking flexibility in handling cases of varying severity. While these methods can identify some risk in typical high-risk patients, excessive testing can lead to unnecessary examinations for low-risk patients, consuming limited resources and time. Conversely, in complex cases with atypical symptoms, insufficient initial evidence may delay the identification of high-risk signals, increasing the likelihood of missed diagnoses. This rigid process results in an imbalance in resource allocation, making it difficult for healthcare professionals in high-traffic environments to simultaneously prioritize the rapid admission of low-risk patients and thorough screening for potentially high-risk cases.

[0004] The core technical challenge lies in how to achieve tiered and progressive utilization of clinical data during the assessment process, combined with a timely termination mechanism. In practice, collecting all information from all patients from the outset, including highly specific biomarkers, significantly prolongs the overall waiting time, preventing patients truly requiring urgent intervention from accessing resources promptly. Conversely, relying solely on initial basic information for rapid assessment may miss crucial risk clues in cases with ambiguous symptoms, leading to repeated supplementary examinations and a fragmented and inefficient process. This inherent contradiction between early warning and avoiding excessive testing among patients at different risk levels directly impacts the accuracy and smoothness of triage.

[0005] Therefore, how to dynamically control the depth of assessment based on individual patient conditions in emergency triage, so as to be able to issue early warnings and terminate low-risk pathways as soon as possible when typical high-risk cases appear, and to introduce more evidence in an orderly manner for complex cases without drawing premature conclusions, has become a key issue in improving the efficiency and safety of diagnosis and treatment. Summary of the Invention

[0006] This invention takes acute aortic dissection, the most high-risk disease in emergency medicine, as an example, and provides a diagnostic and predictive method for acute aortic dissection in a multicenter setting, mainly including:

[0007] The agent system acquires the patient's first-level clinical data and converts it into narrative clinical description information. The clinical description information ends with a diagnostic question and requires the selection of a suspected or non-target disease.

[0008] The large language model performs semantic understanding on the clinical description information and generates a first-level risk assessment, selecting the target disease as a suspected or non-target disease.

[0009] If the first-level risk assessment indicates a suspected target disease, the agent system terminates subsequent data input and outputs a target disease risk warning; otherwise, the agent system acquires the patient's second-level clinical data and incorporates it into the clinical description information.

[0010] The large language model generates a second-level risk assessment based on the updated clinical description information, and selects a suspected or non-target disease. If the target disease is suspected, the agent system terminates subsequent data input and outputs a target disease risk warning; otherwise, the agent system obtains the patient's third-level clinical data and incorporates it into the clinical description information.

[0011] The large-scale language model generates a third-layer risk assessment based on the final clinical description information, selects the target disease, suspected or non-target disease, and outputs the final prediction result.

[0012] Preferably, as one possible implementation, the agent system acquires the patient's first-level clinical data and converts it into narrative clinical description information, including:

[0013] The agent system extracts patient demographic data and symptom data from electronic medical records. The demographic data includes age and gender, and the symptom data includes onset time, whether the symptoms are sudden, blood pressure, and past medical history.

[0014] The agent system organizes the demographic and symptom data into narrative clinical description information. The clinical description information integrates the patient's basic characteristics, medical history, and symptom presentation, and ends with a diagnostic question, requiring the selection of a target disease, a suspected disease, or a non-target disease.

[0015] The large language model receives the clinical description information and performs semantic understanding, which determines the baseline risk probability and generates a first-level risk assessment based on cueing engineering techniques.

[0016] Preferably, as one possible implementation, the large language model generates a second-layer risk assessment based on the updated clinical description information, including:

[0017] The agent system acquires patient hematological parameter data, including white blood cell count and red blood cell count;

[0018] The agent system incorporates the hematological parameter data into the clinical description information to form an updated clinical description information. The updated clinical description information integrates the first-level clinical data and hematological parameter data and ends with a diagnostic question, requiring the selection of a suspected or non-target disease.

[0019] The large language model performs semantic understanding on the updated clinical description information and generates a second-layer risk assessment, selecting the target disease as a suspected or non-target disease.

[0020] Preferably, as one possible implementation, the large language model generates a third-level risk assessment based on the final clinical description information, including:

[0021] After the agent system determines that the patient is suspected of having a non-target disease in the second-level risk assessment, it acquires the patient's key biomarker data, which includes at least a quantitative value of D-dimer. The quantitative value of D-dimer is then aligned with the hematological parameters in the acquired first-level and second-level clinical data and verified with timestamps to form a standardized final clinical description.

[0022] The large language model receives the final clinical description, performs multi-dimensional semantic association to extract the absolute value of the D-dimer quantitative value, and further performs cross-layer semantic association with "whether the symptoms are sudden" in the first layer data and "white blood cell count" in the second layer data to construct an enhanced semantic network that includes temporal evolution and physiological indicator synergy.

[0023] Based on the enhanced semantic network, the large language model generates a third-layer risk assessment result; the third-layer risk assessment result is a binary prediction result of suspected target disease / suspected non-target disease and is accompanied by a final decision confidence score;

[0024] The agent system outputs a final prediction result, which includes the binary prediction result, the final decision confidence score, and target examination recommendation information based on the binary prediction result; wherein, if the decision is "suspected target disease", the target examination recommendation information is to recommend emergency computed tomography angiography.

[0025] Preferably, as one possible implementation, based on the enhanced semantic network, the large language model generates a third-layer risk assessment result, specifically including:

[0026] The agent system incorporates a particle swarm optimization module, the optimization objective of which is to maximize the overall prediction strength of the diagnostic prediction; the overall prediction strength is dynamically weighted and calculated from the following strength parameters:

[0027] The intensity parameter includes "whether the symptoms are sudden" in the first layer of data, which describes whether the patient's vital signs and symptoms are sudden.

[0028] In the second layer of data, "white blood cell count" is used to describe an abnormally high number of cells that exceeds the normal range by a factor of 5.

[0029] In the third layer of data, the "D-dimer quantitative value" is used to describe the intensity exceeding the standard threshold by a factor of 5.

[0030] The particle swarm optimization (PSO) module takes the aforementioned strength parameters and the final decision confidence score as input vectors, and iteratively searches for the optimal weight combination to maximize the particle swarm fitness function value, thereby guiding the large language model to output a third-layer risk assessment result with the optimal comprehensive prediction strength.

[0031] Preferably, as one possible implementation, the comprehensive predicted intensity is quantized into a scalar value IDS, which is calculated using the following vectorized formula:

[0032] ;

[0033] in: This represents the intensity value of the acute state feature extracted from the first layer of clinical data. Let m be the total number of symptom features, and let m be the weight corresponding to each feature. This represents the intensity value of the j-th inflammatory biomarker extracted from the second-level clinical data. The corresponding weights are given by n, where n is the total number of inflammatory markers. This represents the intensity value of the k-th key biomarker extracted from the third-level clinical data. Let p be the total number of key markers, and its corresponding weight.

[0034] C represents the final decision confidence score, and δ represents its weight; α, β, γ, and δ are hierarchical adjustment coefficients, satisfying α + β + γ + δ = 1. It should be noted that the specific values ​​of α, β, γ, and δ are determined by the particle swarm optimization module through iterative optimization using historical training data.

[0035] Preferably, as one possible implementation, the particle swarm optimization module determines the weights α, β, γ, δ and , , The process involves introducing multi-center data difference compensation processing, specifically including:

[0036] The module extracts complete three-tier clinical data and final diagnostic labels of confirmed positive and negative cases from the electronic medical record historical databases of each participating center to form a training set;

[0037] For each case, a set of particles is initialized, and the position vector of each particle encodes all the weight parameters (α, β, γ, δ) in the formula. , , A set of candidate values ​​for ).

[0038] Calculate the fitness value for each particle, whereby the fitness function is defined as:

[0039] Wherein, AUC is the area under the curve obtained by using the set of weight parameters to calculate IDS on all training set cases and using it for diagnostic prediction; λ represents the absolute difference in the mean IDS of similar cases between two centers A and B with different case sources; λ is the difference penalty coefficient.

[0040] By iteratively updating the velocity and position of each particle, we search for a fitness function that optimizes the fitness function. Maximize the globally optimal solution.

[0041] Preferably, as one possible implementation, the agent system terminates subsequent data input and outputs a target disease risk warning, including:

[0042] The agent system stores the first-level risk assessment or the second-level risk assessment as historical case data;

[0043] The agent system generates interpretable output based on the historical case data through contextual reasoning, and the interpretable output includes the intermediate risk assessment results at each decision node;

[0044] The agent system controls the data input process through an adaptive multi-turn dialogue mechanism, which determines whether to advance to the next level of clinical data based on the risk assessment results.

[0045] Preferably, as one possible implementation, the large language model receives the clinical description information and performs semantic understanding, including: the prompting engineering technology sets the temperature parameter to zero; the large language model receives the clinical description information through an application programming interface; the large language model determines the semantic association of each clinical data in the clinical description information and generates a first-level risk assessment, selecting a suspected or non-target disease.

[0046] Preferably, as one possible implementation, the large language model performs semantic understanding on the updated clinical description information, including: the large language model integrates the hematological parameter data with the first-layer clinical data to determine biological relevance; the large language model generates a second-layer risk assessment based on the updated clinical description information, and selects suspected or non-target diseases.

[0047] The agent system acquires patient clinical data, including: the clinical data comes from electronic medical records and is detected within a preset time after the patient is admitted to the hospital; the agent system divides the clinical data into first-level clinical data, second-level clinical data and third-level clinical data.

[0048] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0049] This invention discloses a diagnostic prediction method for acute aortic dissection in a multi-center environment. Addressing the unique business scenario of rapid patient triage and optimized resource allocation in emergency settings, it innovatively solves the balance between early identification of high-risk patients and excessive testing of low-risk patients. In high-traffic emergency environments, this invention achieves orderly risk assessment from basic information to specific biomarkers through a hierarchical and progressive clinical data evaluation mechanism combined with the semantic understanding capabilities of a large-scale language model. Specifically, it can issue early warnings for typical high-risk cases at the first level and avoid unnecessary testing through an early termination mechanism. Simultaneously, it gradually introduces more evidence for complex cases to ensure diagnostic accuracy. This invention also dynamically controls data input through an adaptive multi-turn dialogue mechanism, optimizes the process in real time, and supports batch processing and risk ranking, helping medical staff prioritize critical cases. Ultimately, this invention significantly improves the efficiency and safety of emergency patient care, reduces the rate of missed diagnoses and medical costs, demonstrating its high applicability and technological value in emergency scenarios. Attached Figure Description

[0050] Figure 1 This is a flowchart of a diagnostic and prediction method for acute aortic dissection in a multicenter environment according to the present invention.

[0051] Figure 2 This is a schematic diagram of a diagnostic and prediction method for acute aortic dissection in a multicenter environment according to the present invention;

[0052] Figure 3 This is another schematic diagram of a diagnostic and prediction method for acute aortic dissection in a multicenter environment according to the present invention;

[0053] Figure 4 This is another schematic diagram of a diagnostic and prediction method for acute aortic dissection in a multicenter environment according to the present invention. Detailed Implementation

[0054] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0055] Example 1

[0056] This embodiment provides a method for the diagnosis and prediction of acute aortic dissection in a multicenter setting, which may specifically include:

[0057] The agent system acquires the patient's first-level clinical data and converts it into narrative clinical description information. The clinical description information ends with a diagnostic question and requires the selection of a suspected or non-target disease.

[0058] The large language model performs semantic understanding on the clinical description information and generates a first-level risk assessment, selecting the target disease as a suspected or non-target disease.

[0059] If the first-level risk assessment indicates a suspected target disease, the agent system terminates subsequent data input and outputs a target disease risk warning; otherwise, the agent system acquires the patient's second-level clinical data and incorporates it into the clinical description information.

[0060] The large language model generates a second-level risk assessment based on the updated clinical description information, and selects a suspected or non-target disease. If the target disease is suspected, the agent system terminates subsequent data input and outputs a target disease risk warning; otherwise, the agent system obtains the patient's third-level clinical data and incorporates it into the clinical description information.

[0061] The large-scale language model generates a third-layer risk assessment based on the final clinical description information, selects the suspected or non-target disease, and outputs the final prediction result. The execution order of the above process can be changed, which will not be elaborated further.

[0062] In the above embodiments, the present invention provides a method for the diagnosis and prediction of acute aortic dissection in a multicenter environment. Figure 1 This is a schematic diagram of the overall process of the target disease risk prediction method provided in the embodiments of the present invention, such as... Figure 1 As shown, the method includes the following steps. Step S1: The agent system acquires the patient's first-level clinical data and converts it into narrative clinical description information, which ends with a diagnostic question, requiring the selection of a suspected or non-target disease.

[0063] Specifically, the agent system first extracts the first-level clinical data from the patient's electronic medical record system. This first-level clinical data mainly includes the patient's demographic and symptom data. The demographic data includes basic information such as age and gender, while the symptom data covers the onset time, whether the symptoms were sudden, blood pressure, and past medical history. This data is usually available shortly after the patient's admission, making it highly accessible and fundamental.

[0064] In one embodiment, the agent system accesses electronic medical records in real time through an interface with the hospital information system. Once a patient's registration is complete, the system automatically extracts relevant fields. For example, for a 55-year-old male patient, the system obtains information such as age 55, gender male, onset time 2 hours prior to visit, sudden onset, and presence of chest constricting pain. The first layer of data, "whether the symptoms are sudden," describes whether the patient's vital signs and symptoms are sudden (e.g., acute aortic dissection patients are in critical condition with very short emergency treatment time, so time is precious; in this case, determining whether it is a sudden condition can be based on the doctor's experience and a score, for example, a score of 8 indicates a sudden condition). Accompanying symptoms include sweating and nausea, blood pressure of 160 / 100 mmHg, and past medical history including hypertension and smoking history. The agent system organizes the extracted data into a coherent narrative clinical description. Narrative clinical descriptions, written in natural language paragraphs, integrate basic patient characteristics, medical history, and symptoms, aligning with the habits of clinicians writing medical records. Specifically, the clinical description first states the patient's basic information, followed by the chief complaint and present illness, then lists relevant physical signs and past medical history, and concludes with a clear diagnostic question, requiring a large language model to select between suspected and non-target diseases.

[0065] In step S2, the large language model performs semantic understanding on the clinical description information and generates a first-level risk assessment, selecting a suspected or non-target disease. Specifically, after receiving the narrative clinical description information transmitted by the agent system, the large language model performs semantic understanding through prompting engineering techniques. In one embodiment, the prompting engineering technique sets the temperature parameter to zero to ensure that the model output is deterministic and repeatable, avoiding assessment fluctuations caused by randomness.

[0066] The large-scale language model first performs semantic association analysis on the various clinical elements in the clinical description information. For example, the model can identify the accompaniment relationship between "sudden onset of squeezing chest pain" and "sweating and nausea," as well as the synergistic risk effects with "age, gender, history of hypertension, and smoking history." Through pre-trained medical knowledge, the model can quickly determine the baseline risk probability without relying on an external knowledge base. The large-scale language model receives the clinical description information through an application programming interface (API) and processes it according to pre-defined system prompts. The system prompts explicitly require the model to strictly base its reasoning on the provided clinical information, without introducing additional assumptions, and the final output format is an explicit choice: suspected target disease or non-target disease. For example, for the aforementioned clinical description information of a 55-year-old male patient, the large-scale language model might analyze underlying diseases (such as hypertension), high-risk age group, male gender, and the coexistence of multiple risk factors, thereby generating a first-level risk assessment of suspected target disease. In another embodiment, if the patient is a 35-year-old female who experiences mild chest tightness without obvious cause, without radiating pain, accompanied by sweating and nausea, has normal blood pressure, and no risk factors for coronary heart disease, the model may assess it as a non-target disease, thereby avoiding unnecessary in-depth examinations for low-risk patients. Step S3: If the first-level risk assessment indicates a suspected target disease, the agent system terminates subsequent data input and outputs a target disease risk warning; otherwise, the agent system acquires the patient's second-level clinical data and incorporates it into the clinical description information. Specifically, the agent system performs branch judgments based on the first-level risk assessment results returned by the large language model. If the assessment result indicates a suspected target disease, the agent system immediately terminates the subsequent clinical data acquisition process, directly generating and outputting a target disease risk warning and corresponding examination suggestions. This allows for early warnings in high-risk patients, shortening decision-making time and improving clinical response efficiency. It should be noted that this early termination mechanism helps optimize the allocation of medical resources, avoids unnecessary waiting for high-risk patients, and reduces the probability of missed diagnoses. Conversely, if the first-level risk assessment indicates a non-target disease, the agent system will continue to acquire the patient's second-level clinical data. The second layer of clinical data primarily consists of hematological parameters, including routine blood indicators such as white blood cell count and red blood cell count. These indicators are usually available during routine blood tests upon patient admission and are highly timely. The agent system integrates the newly acquired hematological parameter data into the existing narrative clinical description information, forming an updated clinical description. For example, adding the following to the original description: "Hematological examination shows a white blood cell count of 12.5 × 10⁻⁶." 9 / L, red blood cell count 4.8×10 12 / L. Maintaining a format ending with a diagnostic question, the large language model is required to reassess whether the target disease is suspected. In one embodiment, for patients assessed as having a non-target disease in the first layer, if the second layer of hematological parameters shows a significantly elevated white blood cell count, it may indicate an infectious disease, further supporting the non-target disease judgment; if the white blood cell count is normal but accompanied by other abnormalities, it may be necessary to proceed to the next layer of assessment. Through this hierarchical and progressive mechanism, the agent system achieves adaptive data input control, dynamically deciding whether to advance to the next layer of clinical data collection based on the risk assessment results, thereby ensuring diagnostic accuracy while reducing unnecessary examinations for low-risk patients.

[0067] In one possible implementation, the agent system interacts with clinical staff through an adaptive multi-turn dialogue mechanism. When second-level data is needed, the system prompts the nursing station or laboratory to provide relevant test results and monitors the result return status in real time to ensure that the process is not excessively prolonged due to waiting. Furthermore, the agent system stores the first-level risk assessment results and corresponding clinical descriptions as historical case data for subsequent contextual reasoning and continuous model optimization. This historical data contains intermediate assessment results at each decision node, providing a traceable basis for clinical review and system improvement.

[0068] Through the implementation of steps S1 to S3 above, this invention can quickly complete preliminary risk screening based on the most basic clinical information in the early stages of patient visits, immediately alerting patients with high risk and systematically introducing more evidence when the risk is low, thereby achieving accurate and efficient prediction of the target disease risk. In another embodiment, for high-traffic scenarios in emergency departments, the agent system can process the first-level data of multiple patients in parallel. When multiple patients are simultaneously assessed as suspected of having the target disease, the system will sort and output the data according to the urgency of the risk alerts, helping medical staff to prioritize the most critical cases. The transition from the first to the second level fully embodies the idea of ​​stratified diagnosis, using the universal availability of demographic and symptom data as an entry barrier to filter out most low-risk cases, and only allocating more testing resources to cases that require further identification, thereby reducing overall medical costs and improving diagnostic efficiency.

[0069] In step S4, the large language model generates a second-level risk assessment based on the updated clinical description information, selecting either a suspected or non-target disease. If the target disease is suspected, the agent system terminates subsequent data input and outputs a risk warning for the target disease; otherwise, the agent system obtains the patient's third-level clinical data and incorporates it into the clinical description information.

[0070] Specifically, after the first-level risk assessment identifies the disease as non-target and the updated clinical description information has been incorporated into the second-level clinical data, the large language model re-receives this updated clinical description information for a deeper semantic understanding. The second-level clinical data mainly consists of hematological parameters, which are readily available during routine admission examinations and can provide important biological evidence for the differential diagnosis of the target disease.

[0071] Large-scale language models generate a second-layer risk assessment by performing holistic semantic analysis of updated clinical descriptions, combined with the correlation between hematological parameters and symptoms, and demographic data. The model focuses on whether hematological parameters support pathological states potentially related to the target disease, such as inflammatory responses or anemia. For example, a significantly elevated white blood cell count may indicate an increased likelihood of infectious diseases, thus maintaining or reinforcing the assessment of non-target diseases; conversely, a normal white blood cell count accompanied by other high-risk characteristics may still prompt the model to remain vigilant and continue assessing potential cardiovascular event risks. It is worth noting that incorporating hematological parameters into the second-layer assessment significantly improves the ability to differentiate atypical cases, avoiding misdiagnosis or missed diagnosis based solely on subjective symptoms.

[0072] If the second-level risk assessment indicates a suspected target disease, the agent system immediately terminates the acquisition of subsequent third-level data and directly outputs a risk warning and examination recommendations for the target disease. This mechanism, which terminates at the intermediate level, can issue timely alerts when sufficient evidence accumulates to support a suspected diagnosis, helping clinicians quickly enter the targeted treatment process.

[0073] In step S5, the large language model generates a third-level risk assessment based on the final clinical description information, selects the target disease, suspected or non-target disease, and outputs the final prediction result.

[0074] Specifically, when the second-layer risk assessment still indicates a non-target disease, the agent system has already incorporated the third-layer clinical data to form the final clinical description information. At this point, the large-scale language model receives this final clinical description information and performs the most comprehensive semantic understanding and risk assessment. The third-layer clinical data consists of key biomarker data, which can provide highly specific diagnostic support for the target disease. In one embodiment, the final clinical description information integrates demographic data, symptom data, hematological parameters, and key biomarker values ​​to form a complete narrative chain. For example: A 55-year-old male patient presents with sudden chest pain accompanied by sweating and nausea, elevated blood pressure, a history of hypertension and smoking, normal white blood cell count in routine blood tests, and a D-dimer quantitative value of 850 ng / mL. Please determine whether the target disease is suspected based on all information. The large-scale language model generates a third-layer risk assessment by deeply analyzing the final clinical description information and integrating the three layers of evidence, directly linking it to the final prediction result. Regardless of whether the result is suspected or not, the system will output a clear prediction result and targeted examination suggestions.

[0075] If the third-level risk assessment indicates a suspected target disease, a high-risk warning is issued, and imaging examinations, such as coronary CT or interventional angiography, are recommended. If it indicates a non-target disease, a low-risk conclusion is issued, and further investigations for alternative diagnoses are suggested. It should be noted that the third-level assessment, as the final checkpoint, can minimize missed diagnoses and avoid unnecessary invasive examinations for low-risk patients.

[0076] The specific operation process of the embodiments of the present invention will be described in detail below: See Figure 2 The specific operation process of step S2 above is as follows:

[0077] Step S21: The agent system extracts patient demographic data and symptom data from the electronic medical record. The demographic data includes age and gender, and the symptom data includes onset time, whether the symptoms are sudden, blood pressure, and past medical history.

[0078] Specifically, the agent system extracts predefined fields from the electronic medical record database using a structured query language. Demographic data is obtained directly from the patient's basic information table, while symptom data is parsed from the chief complaint, present illness history, and physical examination records. In one embodiment, the system employs natural language processing technology to assist in extracting symptom descriptions from unstructured text, ensuring information integrity.

[0079] In step S22, the agent system organizes the demographic and symptom data into narrative clinical description information. This clinical description information integrates the patient's basic characteristics, medical history, and symptom presentation, ending with a diagnostic question that prompts the user to select a suspected or non-target disease. Specifically, the agent system follows clinical record writing standards, beginning with a statement of age, gender, and chief complaint, followed by a paragraph describing the present illness, past medical history, and physical examination, ending with a standard diagnostic question. This structure facilitates the simulation of clinical thought processes using large-scale language models.

[0080] Step S23: The large language model receives the clinical description information and performs semantic understanding. This semantic understanding, based on cue engineering techniques, determines the baseline risk probability and generates a first-level risk assessment. Cue engineering refers to designing input text (cue words) for the LLM to precisely guide its output (format, content, role, etc.). For example, this template might include: Role setting: "You are an experienced emergency room physician." Task definition: "Your task is to read the following patient summary and assess the probability that they have [target disease]." Input format: "Summary: [Insert previously generated narrative clinical description information here]." Output instructions: "Please output in the following format: 1. Baseline risk probability: (A value between 0-100%, representing the probability under the initial clinical impression). 2. Main supporting reasons: (List 2-3 key pieces of information in the summary that most support this judgment). 3. Main opposing reasons: (List 1-2 pieces of information in the summary that do not support or are warning). 4. First-level assessment conclusion: (Based on the above analysis, select: A. Target disease suspected / B. Target disease not suspected)." It should be noted that the final decision confidence score (C) can be the probability value of the corresponding "suspected" or "non-suspected" label output by the LLM through its internal softmax layer when generating text, which will not be elaborated further.

[0081] In one embodiment, when typical symptoms and multiple risk factors are present in the clinical description information, the model will quickly tend to be suspected of being the target disease; when the symptoms are atypical or there are alternative explanations, it will tend to be a non-target disease.

[0082] Step S211: The prompting engineering technology sets the temperature parameter to zero. Specifically, setting the temperature parameter to zero allows the model to use deterministic sampling during generation, avoiding different evaluation results from the same input due to randomness, thereby ensuring the consistency of clinical decisions. Step S212: The large language model receives the clinical description information through an application programming interface (API). Specifically, the agent system sends the clinical description information text to the large language model server through a secure API call and receives the returned structured evaluation results. Step S213: The large language model determines the semantic relationships between the clinical data in the clinical description information and generates a first-level risk assessment, selecting suspected or non-target diseases. Specifically, the model internally captures the correlation between symptoms and their synergistic effects with risk factors through an attention mechanism, ultimately forming a comprehensive risk judgment.

[0083] In one embodiment, the model significantly increases the suspected disease probability for patients with chest pain accompanied by sweating, nausea, and elevated blood pressure; while maintaining a low-risk assessment for younger patients presenting only with transient, needle-like chest pain. For example, the output might be: "The patient's current evidence supports a high suspicion of a certain disease; it is recommended to immediately complete troponin series testing and prepare for emergency interventional treatment." Conversely, if the second-level risk assessment still indicates a non-target disease, the agent system continues, acquiring the patient's third-level clinical data and incorporating clinical descriptive information to provide more specific biomarker evidence for the final assessment.

[0084] See Figure 3 The specific operation process of step S3 above is as follows:

[0085] In one embodiment, in step S31, the agent system acquires the patient's hematological parameter data, including white blood cell count and red blood cell count. Specifically, the agent system automatically extracts the specific values ​​of white blood cell count and red blood cell count after the blood routine report is generated, through an interface with the hospital's laboratory information system. These parameters are usually reported within 1 to 2 hours after the patient's admission, exhibiting high timeliness. When the blood routine results are abnormal, the system prioritizes marking them and quickly incorporates them into the clinical description information, avoiding delays in the assessment process due to waiting for reports. In step S32, the agent system incorporates the hematological parameter data into the clinical description information to form an updated clinical description information. The updated clinical description information integrates the first-level clinical data and hematological parameter data and ends with a diagnostic question, requiring the selection of a suspected or non-target disease. Specifically, the agent system naturally integrates the hematological parameters into the original narrative description according to chronological order and clinical logic, usually placing it after the physical examination section to ensure the coherence and readability of the description. The updated clinical description information still ends with a clear diagnostic question, facilitating standardized reasoning by large language models. For example, the additional information could be: Routine blood tests upon admission showed a slightly elevated white blood cell count and a normal red blood cell count. Please reassess whether the target disease is suspected based on the above information. Step S33: The large language model performs semantic understanding on the updated clinical description information and generates a second-layer risk assessment, selecting whether the target disease is suspected or not. Specifically, after receiving the updated clinical description information, the large language model re-analyzes the semantic relationships of all included information, particularly the degree of matching between hematological parameters and symptom presentation. The model utilizes medical knowledge accumulated during pre-training to determine whether the current evidence is sufficient to support a suspected diagnosis of the target disease.

[0086] In one embodiment, when hematological parameters show a significant increase in white blood cell count and symptoms are more consistent with infection, the model will firmly select a non-target disease; when hematological parameters are normal and symptoms are highly typical, even if the first layer is not suspected, the model may switch to suspected target disease in the second layer. Step S311: The large language model integrates the hematological parameter data with the first-layer clinical data to determine biological relevance. Specifically, the model uses an attention mechanism to focus on the interaction between hematological parameters and age, gender, and symptom characteristics.

[0087] In step S312, the large-scale language model generates a second-level risk assessment based on the updated clinical description information, selecting suspected or non-target diseases. Specifically, after comprehensive analysis, the model outputs the second-level risk assessment results in a strict binary selection format, ensuring that the assessment conclusion is clear and unambiguous. Through the implementation of the above-mentioned second-level assessment, this invention can further introduce routine laboratory evidence after initial screening, achieving refined adjustment of risk and providing a reliable basis for whether to proceed with the most specific biomarker detection.

[0088] In one embodiment, in step S41, the agent system acquires key biomarker data of the patient, including D-dimer quantitative values. Specifically, after the second-level assessment identifies the disease as non-target, the agent system immediately triggers the extraction of key biomarker test results. D-dimer quantitative values, as important indicators for excluding or supporting certain acute events, are typically obtainable quickly in emergency rapid testing channels. When D-dimer levels are significantly elevated, the system prioritizes and marks them as high-concern items for focused model analysis. In step S42, the agent system incorporates the key biomarker data into the clinical description information to form final clinical description information. This final clinical description information integrates the first two layers of clinical data and ends with a diagnostic question, requiring the selection of a suspected or non-target disease. In step S43, a large-scale language model performs semantic understanding on the final clinical description information and generates a third-level risk assessment, selecting a suspected or non-target disease. Specifically, after receiving the final clinical description information, the model pays particular attention to the degree of matching between the key biomarkers and the aforementioned evidence. For example, elevated D-dimer levels in patients with chest pain may indicate multiple diseases, and the model needs to combine symptom typicality with hematological parameters for comprehensive differentiation. When D-dimer is normal and symptoms are atypical, the model will firmly select a non-target disease; when D-dimer is elevated but symptoms highly match the target disease, the model will select a suspected target disease. In step S44, the agent system outputs the final prediction result, and the examination recommendations suggest imaging examinations for suspected target diseases. Specifically, the agent system generates a structured output report based on the third-level risk assessment results, including the final prediction conclusion, risk stratification, and specific examination recommendations. For suspected cases of the target disease, coronary artery imaging is given priority to clarify the diagnosis. For example, the output could be: the final assessment is a suspected target disease, and immediate coronary CT angiography or emergency coronary angiography is recommended. Through the implementation of the third-level assessment, this invention realizes a complete diagnostic chain of progressively advancing evidence, ensuring that the most specific evidence is introduced when necessary, ultimately providing highly reliable prediction results.

[0089] See Figure 4 The specific operation process of step S5 above is as follows:

[0090] In one embodiment, step S5, where the large language model generates a third-layer risk assessment based on the final clinical description information, includes:

[0091] S51: After the agent system determines that the patient is suspected of having a non-target disease in the second-level risk assessment, it acquires the patient's key biomarker data, wherein the key biomarker data includes at least a quantitative value of D-dimer; the quantitative value of D-dimer is aligned with the hematological parameters in the acquired first-level clinical data and second-level clinical data and timestamp verification is performed to form a standardized final clinical description;

[0092] S52: The large language model receives the final clinical description, performs multi-dimensional semantic association to extract the absolute value of the D-dimer quantitative value, and further performs cross-layer semantic association with "whether the symptoms are sudden" in the first layer data and "white blood cell count" in the second layer data to construct an enhanced semantic network that includes time evolution and physiological indicator synergy.

[0093] S53: Based on the enhanced semantic network, the large language model generates a third-layer risk assessment result; the third-layer risk assessment result is a binary prediction result of suspected target disease / suspected non-target disease and is accompanied by a final decision confidence score;

[0094] S54: The agent system outputs the final prediction result, which includes the binary prediction result, the final decision confidence score, and the target examination suggestion information based on the binary prediction result; wherein, if the decision is "suspected target disease", the target examination suggestion information is to recommend emergency computed tomography angiography.

[0095] Preferably, as one possible implementation, based on the enhanced semantic network, the large language model generates a third-layer risk assessment result, specifically including:

[0096] The agent system incorporates a particle swarm optimization module, the optimization objective of which is to maximize the overall prediction strength of the diagnostic prediction; the overall prediction strength is dynamically weighted and calculated from the following strength parameters:

[0097] The intensity parameter includes "whether the symptoms are sudden" in the first layer of data, which describes whether the patient's vital signs and symptoms are sudden.

[0098] In the second layer of data, "white blood cell count" is used to describe an abnormally high number of cells that exceeds the normal range by a factor of 5.

[0099] In the third layer of data, the "D-dimer quantitative value" is used to describe the intensity exceeding the standard threshold by a factor of 5.

[0100] The particle swarm optimization (PSO) module takes the aforementioned strength parameters and the final decision confidence score as input vectors, and iteratively searches for the optimal weight combination to maximize the particle swarm fitness function value, thereby guiding the large language model to output a third-layer risk assessment result with the optimal comprehensive prediction strength.

[0101] Preferably, as one possible implementation, the comprehensive predicted intensity is quantized into a scalar value IDS, which is calculated using the following vectorized formula:

[0102] ;

[0103] in: This represents the intensity value of the sudden state feature extracted from the first layer of clinical data (this value can also be an assessment value that can be quickly determined manually through expert consultation). Let m be the total number of symptom features, and let m be the weight corresponding to each feature. This represents the intensity value of the j-th inflammatory biomarker extracted from the second-level clinical data. The corresponding weights are given by n, where n is the total number of inflammatory markers. This represents the intensity value of the k-th key biomarker extracted from the third-level clinical data. Let p be the total number of key markers, and its corresponding weight.

[0104] C represents the final decision confidence score, and δ represents its weight; α, β, γ, and δ are hierarchical adjustment coefficients, satisfying α + β + γ + δ = 1. It should be noted that the specific values ​​of α, β, γ, and δ are determined by the particle swarm optimization module through iterative optimization using historical training data.

[0105] It should be noted that the sudden state characteristic intensity value ( The calculation of SF is as follows: For each acute condition feature (such as 'whether the symptoms are acute'), its qualitative description or raw score is mapped to an intensity value in the range of 0 to 1. For example, if the raw data is a score of 1-10 based on the doctor's experience, it is obtained by linearly normalizing SF = (raw score - 1) / 9.

[0106] Regarding the intensity values ​​of inflammatory biomarkers ( The calculation of white blood cell count, or other routine hematological indicators, first involves calculating the degree of deviation based on the ratio of the count to the upper limit of normal. The formula is: = min(detection value / ULNj, upper cutoff value). The upper cutoff value (e.g., set to 5) is used to prevent excessive influence of individual extreme values ​​on the model, ensuring numerical stability. ULNj represents the upper limit of the normal value for the j-th inflammatory biomarker;

[0107] Key biomarker intensity values ​​( The calculation of BFk, for example, uses a threshold-based multiple intensity calculation method for key markers such as D-dimer. The formula is: BFk = min(D-dimer quantitative value / diagnostic threshold, upper cutoff value). Here, the 'diagnostic threshold' is a standard value used clinically to aid in the diagnosis of acute aortic dissection (e.g., 500 ng / mL).

[0108] Preferably, as one possible implementation, the particle swarm optimization module determines the weights α, β, γ, δ and , , The process involves introducing multi-center data difference compensation processing, specifically including:

[0109] S71: The module extracts complete three-layer clinical data and final diagnostic labels of confirmed positive and negative cases from the electronic medical record historical database of each participating center to form a training set;

[0110] S72: For each case, initialize a set of particles, where the position vector of each particle encodes all weight parameters (α, β, γ, δ) in the formula. , , A set of candidate values ​​for ).

[0111] S73: Calculate the fitness value for each particle, whereby the fitness function is defined as: ;

[0112] Wherein, AUC is the area under the curve obtained by using the set of weight parameters to calculate IDS on all training set cases and using it for diagnostic prediction; λ represents the absolute difference in the mean IDS of similar cases between two centers A and B with different case sources; λ is the difference penalty coefficient.

[0113] S74: By iteratively updating the velocity and position of each particle, find a fitness function that satisfies the condition. The global optimal solution is maximized, and the set of weight parameters corresponding to this optimal solution is the optimized weight of the formula finally applied to the comprehensive predictive intensity IDS, thereby ensuring that the comprehensive predictive intensity IDS has both high diagnostic efficacy and evaluation consistency in a multi-center environment.

[0114] In the specific technical solution, the following operation is included after step S54: Step S55, the agent system stores the first-level risk assessment or the second-level risk assessment as historical case data. Specifically, regardless of which level the assessment terminates, the agent system will completely store the risk assessment results, corresponding clinical description information, and decision path of that level into a dedicated historical case database. This data includes the specific inputs and model outputs of each decision node, forming a traceable and complete record. Step S56, the agent system performs contextual reasoning based on the historical case data to generate interpretable output, which includes the intermediate risk assessment results at each decision node. Specifically, when outputting the final result, the agent system will also attach an interpretable report, listing the key evidence and intermediate conclusions of each level of assessment. For example, the report may state that the first level was rated as suspected due to typical symptoms, or the second level remained non-suspected due to normal hematological parameters.

[0115] In step S57, the agent system controls the data input process through an adaptive multi-turn dialogue mechanism. This mechanism determines whether to advance to the next level of clinical data based on the risk assessment results. Specifically, the agent system maintains real-time interaction with the healthcare terminal. When the next level of data is needed, a supplementary request automatically pops up; when the assessment indicates a suspected case, a risk alert is immediately pushed to the attending physician's workstation. The dialogue mechanism supports manual intervention; for example, doctors can manually trigger the inclusion of a certain level of data to address special clinical situations. Through the adaptive multi-turn dialogue mechanism, the entire prediction process achieves dynamic adjustment, ensuring both automation and efficiency while retaining clinical flexibility.

[0116] The specific technical solution also includes the following operation: Step S81, the clinical data comes from the electronic medical record and is detected within a preset time after the patient's admission. Specifically, all three layers of clinical data come from the hospital's electronic medical record system and are collected and detected within a specified time after the patient's admission. The first layer of data is available immediately after the consultation, the second layer after the routine blood report, and the third layer after rapid biomarker detection.

[0117] In step S82, the agent system stratifies the clinical data into three layers: a first layer, a second layer, and a third layer. Specifically, the agent system predefines stratification rules: the first layer consists of medical history and vital signs information; the second layer consists of routine hematological indicators; and the third layer consists of key specific biomarkers. This stratification design is based on the strength and ease of obtaining clinical evidence, progressing from easy to difficult. The stratification rules can be fine-tuned according to the specific treatment guidelines of the hospital, but the three-layer structure remains unchanged. Through the implementation of the above entire process, this invention provides a stratified, evidence-driven method for predicting the risk of targeted diseases. Utilizing the powerful semantic understanding capabilities of a large-scale language model, combined with the intelligent data control of the agent system, it achieves an orderly assessment from basic information to specific biomarkers. In clinical practice, for typical high-risk chest pain patients, the system often terminates and issues a warning at the first or second layer, significantly shortening the diagnosis time; for complex cases with atypical symptoms, the system systematically introduces more evidence, ultimately providing an accurate conclusion.

[0118] It should be noted that the entire method does not rely on external knowledge base updates; it is entirely based on medical knowledge embedded in a large language model and real-time clinical data, enabling rapid deployment and application across different medical institutions. Through multi-layered evaluation and early termination mechanisms, this invention maintains high sensitivity while controlling specificity, avoiding overtreatment, and providing interpretable supporting evidence for clinical decision-making, thereby improving the overall efficiency and safety of chest pain patient diagnosis and treatment.

[0119] In one possible implementation, the system also supports batch processing of multiple chest pain patients, automatically sorting and outputting reports according to assessment level and risk level to help emergency departments optimize triage and resource allocation. For example, during peak night shift periods, when multiple patients seek medical attention simultaneously, the system prioritizes pushing reports of suspected cases (level 1) to ensure that high-risk patients receive timely treatment.

[0120] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for diagnosis prediction of acute aortic dissection in a multi-center environment, characterized in that, include: The agent system acquires the patient's first-level clinical data and converts it into narrative clinical description information. The clinical description information ends with a diagnostic question and requires the selection of a suspected or non-target disease. The large language model performs semantic understanding on the clinical description information and generates a first-level risk assessment, selecting the target disease as a suspected or non-target disease. If the first-level risk assessment indicates a suspected target disease, the agent system terminates subsequent data input and outputs a target disease risk warning; otherwise, the agent system acquires the patient's second-level clinical data and incorporates it into the clinical description information. The large language model generates a second-level risk assessment based on the updated clinical description information, and selects a suspected or non-target disease. If the target disease is suspected, the agent system terminates subsequent data input and outputs a target disease risk warning; otherwise, the agent system obtains the patient's third-level clinical data and incorporates it into the clinical description information. The large-scale language model generates a third-layer risk assessment based on the final clinical description information, selects the target disease, suspected or non-target disease, and outputs the final prediction result.

2. The method of claim 1, wherein, The agent system acquires the patient's first-level clinical data and converts it into narrative clinical description information, including: The agent system extracts patient demographic data and symptom data from electronic medical records. The demographic data includes age and gender, and the symptom data includes onset time, whether the symptoms are sudden, blood pressure, and past medical history. The agent system organizes the demographic and symptom data into narrative clinical description information. The clinical description information integrates the patient's basic characteristics, medical history, and symptom presentation, and ends with a diagnostic question, requiring the selection of a target disease, a suspected disease, or a non-target disease. The large language model receives the clinical description information and performs semantic understanding, which determines the baseline risk probability and generates a first-level risk assessment based on cueing engineering techniques.

3. The method as described in claim 1, characterized in that, The large language model generates a second-level risk assessment based on the updated clinical description information, including: The agent system acquires patient hematological parameter data, including white blood cell count and red blood cell count; The agent system incorporates the hematological parameter data into the clinical description information to form an updated clinical description information. The updated clinical description information integrates the first-level clinical data and hematological parameter data and ends with a diagnostic question, requiring the selection of a suspected or non-target disease. The large language model performs semantic understanding on the updated clinical description information and generates a second-layer risk assessment, selecting the target disease as a suspected or non-target disease.

4. The method as described in claim 1, characterized in that, The large language model generates a third-level risk assessment based on the final clinical description information, including: After the agent system determines that the patient is suspected of having a non-target disease in the second-level risk assessment, it acquires the patient's key biomarker data, which includes at least a quantitative value of D-dimer. The quantitative value of D-dimer is then aligned with the hematological parameters in the acquired first-level and second-level clinical data and verified with timestamps to form a standardized final clinical description. The large language model receives the final clinical description, performs multi-dimensional semantic association to extract the absolute value of the D-dimer quantitative value, and further performs cross-layer semantic association with "whether the symptoms are sudden" in the first layer data and "white blood cell count" in the second layer data to construct an enhanced semantic network that includes temporal evolution and physiological indicator synergy. Based on the enhanced semantic network, the large language model generates a third-layer risk assessment result; the third-layer risk assessment result is a binary prediction result of suspected target disease / suspected non-target disease and is accompanied by a final decision confidence score; The agent system outputs a final prediction result, which includes the binary prediction result, the final decision confidence score, and target examination recommendation information based on the binary prediction result; wherein, if the decision is "suspected target disease", the target examination recommendation information is to recommend emergency computed tomography angiography.

5. The method as described in claim 4, characterized in that, Based on the enhanced semantic network, the large language model generates a third-layer risk assessment result, specifically including: The agent system incorporates a particle swarm optimization module, the optimization objective of which is to maximize the overall prediction strength of the diagnostic prediction; the overall prediction strength is dynamically weighted and calculated from the following strength parameters: The intensity parameter includes "whether the symptoms are sudden" in the first layer of data, which describes whether the patient's vital signs and symptoms are sudden. In the second layer of data, "white blood cell count" is used to describe an abnormally high number of cells that exceeds the normal range by a factor of 5. In the third layer of data, the "D-dimer quantitative value" is used to describe the intensity exceeding the standard threshold by a factor of 5. The particle swarm optimization (PSO) module takes the aforementioned strength parameters and the final decision confidence score as input vectors, and iteratively searches for the optimal weight combination to maximize the particle swarm fitness function value, thereby guiding the large language model to output a third-layer risk assessment result with the optimal comprehensive prediction strength.

6. The method as described in claim 5, characterized in that, The comprehensive predicted intensity is quantized into a scalar value IDS, which is calculated using the following vectorized formula: ； in: This represents the intensity value of the acute state feature extracted from the first layer of clinical data. Let m be the total number of symptom features, and its corresponding weight. This represents the intensity value of the j-th inflammatory biomarker extracted from the second-level clinical data. The corresponding weights are given by n, where n is the total number of inflammatory markers. This represents the intensity value of the k-th key biomarker extracted from the third-level clinical data. Let p be the total number of key markers, and its corresponding weight. C is the final decision confidence score, and δ is its weight; α, β, γ, δ are the hierarchical adjustment coefficients, and satisfy α+β+γ+δ= 1.

7. The method as described in claim 6, characterized in that, The particle swarm optimization module determines the weights α, β, γ, δ, and , , The process involves introducing multi-center data difference compensation processing, specifically including: The module extracts complete three-tier clinical data and final diagnostic labels of confirmed positive and negative cases from the electronic medical record historical databases of each participating center to form a training set; For each case, a set of particles is initialized, and the position vector of each particle encodes all the weight parameters (α, β, γ, δ) in the formula. , , A set of candidate values ​​for ). Calculate the fitness value for each particle, whereby the fitness function is defined as: ; Wherein, AUC is the area under the curve obtained by using the set of weight parameters to calculate IDS on all training set cases and using it for diagnostic prediction; λ represents the absolute difference in the mean IDS of similar cases between two centers A and B with different case sources; λ is the difference penalty coefficient. By iteratively updating the velocity and position of each particle, we search for a fitness function that optimizes the fitness function. Maximize the globally optimal solution.

8. The method as described in claim 1, characterized in that, The agent system terminates subsequent data input and outputs a target disease risk alert, including: The agent system stores the first-level risk assessment or the second-level risk assessment as historical case data; The agent system generates interpretable output based on the historical case data through contextual reasoning, and the interpretable output includes the intermediate risk assessment results at each decision node; The agent system controls the data input process through an adaptive multi-turn dialogue mechanism, which determines whether to advance to the next level of clinical data based on the risk assessment results.

9. The method as described in claim 2, characterized in that, The large language model receives the clinical description information and performs semantic understanding, including: the prompting engineering technology sets the temperature parameter to zero; the large language model receives the clinical description information through an application programming interface; the large language model determines the semantic association of each clinical data in the clinical description information and generates a first-level risk assessment, selecting a suspected or non-target disease.

10. The method as described in claim 3, characterized in that, The large language model performs semantic understanding on the updated clinical description information, including: The large language model integrates the hematological parameter data with the first-level clinical data to determine biological relevance; the large language model generates a second-level risk assessment based on the updated clinical description information, selecting suspected or non-target diseases; The agent system acquires patient clinical data, including: the clinical data comes from electronic medical records and is detected within a preset time after the patient is admitted to the hospital; the agent system divides the clinical data into first-level clinical data, second-level clinical data and third-level clinical data.

Images