A method, device and electronic equipment for predicting risk of cardiac amyloidosis

Abstract

The application provides a heart amyloidosis risk prediction method, device and electronic equipment. The method comprises the following steps: obtaining medical information of a suspected patient; analyzing and processing the medical information by a large language model to obtain auxiliary prediction information containing unsupported point information and triage suggestion results; screening department doctors according to the triage suggestion results; sending the auxiliary prediction information to the terminal equipment of the department doctors according to the triage suggestion results, so as to receive the newly added diagnosis information and the subsequent medical information of the suspected patient fed back by the doctors through the terminal equipment of the department doctors; if the newly added diagnosis information does not contain the exclusion or diagnosis of the heart amyloidosis risk, analyzing and processing the subsequent medical information of the suspected patient by the large language model until the unsupported point information in the newly added diagnosis information no longer prompts the need for additional examination items; wherein the newly added diagnosis information and the auxiliary prediction information are used to exclude or diagnose the heart amyloidosis risk.

Description

Technical Field

[0001] This application relates to the technical fields of medical artificial intelligence, intelligent risk prediction and intelligent triage, and more specifically, to a method, device and electronic device for predicting the risk of cardiac amyloidosis. Background Technology

[0002] Current preliminary prediction processes for common diseases mostly involve a single model call to predict disease risk. However, this cannot be done with a single model call for cardiac amyloidosis. Currently, the process typically involves doctors from relevant departments suspecting a possible risk of cardiac amyloidosis and then convening a multi-outpatient medical team (including cardiology, hematology, and nephrology specialists) to discuss the affected organs. Further examination of the affected organs is then conducted until a diagnosis is confirmed or ruled out. Therefore, convening a multi-outpatient medical team for early risk prediction of rare diseases is less efficient. Summary of the Invention

[0003] The purpose of this application is to provide a method, apparatus, and electronic device for predicting the risk of cardiac amyloidosis, in order to improve the problem of low efficiency in early risk prediction for rare diseases.

[0004] This application provides a method for predicting the risk of cardiac amyloidosis, comprising: acquiring the medical information of a suspected patient; analyzing and processing the medical information using a large language model to obtain auxiliary predictive information including unsupported point information and triage suggestion results; screening departmental doctors based on the triage suggestion results; sending the auxiliary predictive information to the terminal device of the departmental doctor based on the triage suggestion results, so that the departmental doctor's terminal device can receive feedback from the doctor on newly added diagnostic information and subsequent medical information of the suspected patient, wherein the newly added diagnostic information includes the results of supplementary examination items based on the unsupported point information; if the newly added diagnostic information does not include the exclusion or confirmation of the risk of cardiac amyloidosis, then analyzing and processing the subsequent medical information of the suspected patient using a large language model until the unsupported point information in the obtained newly added diagnostic information no longer indicates the need for supplementary examination items; wherein, the newly added diagnostic information and auxiliary predictive information are used to exclude or confirm the risk of cardiac amyloidosis. In implementing the above solution, a large language model is used to automatically analyze patient medical information, intelligently identify potential risk characteristics of cardiac amyloidosis, and automatically generate auxiliary predictive information including information on points of incompatibility and triage suggestions, rather than relying on the doctor's subjective suspicion. This intelligently recommends relevant departments (such as cardiology, hematology, and nephrology) for consultation and pushes the screening results to the corresponding doctors' terminals. The analysis is then continuously optimized based on the doctor's additional diagnostic information, forming a closed-loop feedback mechanism until the risk of cardiac amyloidosis is confirmed or ruled out. This closed-loop feedback mechanism achieves a shift from "doctor's proactive suspicion" to "system-driven proactive screening," effectively improving the problems of missed and misdiagnosed cases in the diagnosis of cardiac amyloidosis caused by the lack of specific symptoms, the need for multidisciplinary collaboration, and long diagnostic cycles, significantly improving the efficiency of early risk prediction for rare diseases.

[0005] Optionally, in this embodiment, the analysis and processing of medical information through a large language model includes: extracting key clinical indicators from the medical information and mapping them to standard terminology clinical indicators; performing an initial search in a treatment guideline knowledge base based on the standard terminology clinical indicators to obtain treatment guideline knowledge blocks; generating an inference path and a precise retrieval question corresponding to the inference path using the large language model based on the treatment guideline knowledge blocks and the medical information; converting the precise retrieval question into a retrieval embedding vector; performing a second search in a vector database using the retrieval embedding vector to obtain an embedding vector knowledge block with a higher relevance to the medical information than the treatment guideline knowledge block, wherein the relevance between the embedding vector knowledge block and the medical information (e.g., the similarity between the embedding vector and the embedding vector converted from the medical information) is greater than the relevance between the treatment guideline knowledge block and the medical information; and inputting the embedding vector knowledge block and the medical information together into the large language model to generate auxiliary prediction information. In implementing the above solution, an initial search of the clinical guidelines knowledge base based on standard terminology and clinical indicators quickly locates the knowledge blocks most relevant to the patient's condition, providing a preliminary reference framework for subsequent precise retrieval and reasoning. This step improves retrieval efficiency and avoids blind searching in a large-scale knowledge base. Furthermore, by generating reasoning paths and precise retrieval questions through a large language model, complex medical information can be transformed into structured retrieval requirements, thereby more accurately locating knowledge blocks highly relevant to the patient's condition. This leverages the natural language understanding and generation capabilities of the large language model, enhancing the targeting and intelligence of the retrieval.

[0006] Optionally, in this embodiment, selecting departmental doctors based on triage recommendations includes: determining the range of selectable doctors based on their appointment times, shift schedules, holiday markers, and peak outpatient index; extracting selectable doctor features from each selectable doctor within the range; calculating the appointment success rate of each selectable doctor using an XGBoost model based on these features and the patient's medical information; and selecting departmental doctors from the range of selectable doctors based on the appointment success rate. In the implementation of the above scheme, by extracting features from the selectable doctor information and combining them with the XGBoost model to calculate the appointment success rate, the powerful feature processing and prediction capabilities of machine learning algorithms are utilized. This allows for the uncovering of the complex nonlinear relationship between doctor features and appointment success rate, significantly improving the accuracy of doctor recommendations compared to traditional manual experience-based judgment or simple rule matching. Furthermore, by selecting based on the calculated appointment success rate, the matching degree of doctors is quantified and ranked, ensuring that the final departmental doctor recommendation results have objective data support, reducing subjective interference, and simultaneously improving patient appointment success rate and satisfaction.

[0007] Optionally, in this embodiment, after extracting the features of each available doctor from the range of available doctors, the method further includes: adding pseudo-sample features with random rewards to each available doctor; merging the available doctor features and the pseudo-sample features with random rewards to obtain merged doctor features; adding the merged doctor features to the available doctor features; and updating the consultation success rate of each available doctor based on the available doctor features. In the implementation of the above scheme, adding pseudo-sample features with random rewards to each available doctor can effectively alleviate the cold start problem in the recommendation system, because the introduction of random rewards can increase the diversity of data, allowing the model to access a wider range of feature distributions in the early stages of training, thus converging to a better initial state more quickly. Furthermore, by adding the merged doctor features to the available doctor features and updating the consultation success rate, the recommendation weight of doctors can be dynamically optimized, because the introduction of random rewards can balance the relationship between exploration and utilization, allowing the system to recommend high-success-rate doctors while also exploring potential high-value doctors.

[0008] Optionally, in this embodiment, after selecting departmental doctors based on triage recommendations, the method further includes: receiving a consultation request sent by the departmental doctor's terminal device; initiating a multi-agent consultation based on the consultation request to obtain consultation results; and sending the consultation results to the departmental doctor's terminal device. In implementing the above solution, by receiving a consultation request sent by the departmental doctor's terminal device and initiating a multi-agent consultation, cross-departmental and cross-professional collaborative diagnosis can be achieved, thereby improving the accuracy and comprehensiveness of diagnosis. Especially in the diagnosis of complex cases, it can integrate knowledge and experience from multiple professional fields, reducing the limitations of a single doctor's subjective judgment. Furthermore, using a multi-agent system for consultation can significantly improve consultation efficiency because multiple agents can process different diagnostic tasks in parallel, quickly integrate and analyze large amounts of medical data, thereby shortening consultation time and avoiding the time-consuming problem of multiple doctors needing to participate simultaneously in traditional consultations. Furthermore, the introduction of the aforementioned multi-agent system makes the consultation process more intelligent and automated, enabling diagnostic reasoning based on preset rules and algorithms, reducing human error and omissions, and continuously optimizing the diagnostic model based on historical data and cases, thereby improving the system's adaptability and diagnostic accuracy.

[0009] Optionally, in this embodiment, the multiple agents include: a main agent and departmental agents; initiating a multi-agent consultation based on a consultation request includes: parsing the consultation question from the consultation request through the main agent, and determining the departmental agents that need to participate based on the consultation question; obtaining the response results of each departmental agent by having each departmental agent answer the consultation question; and obtaining the consultation result by summarizing the response results of each departmental agent through the main agent. In the implementation of the above scheme, by setting a hierarchical structure of the main agent and departmental agents, intelligent allocation and collaborative processing of consultation questions are achieved. The main agent can automatically parse the consultation request and determine the relevant departmental agents, avoiding the inefficiency and possible omissions of manual allocation, and improving the accuracy and comprehensiveness of the consultation. Furthermore, by summarizing the response results of each departmental agent through the main agent, the professional opinions of different departments can be integrated to form a comprehensive and systematic consultation result. This collaborative working mechanism overcomes the problem of fragmented and difficult-to-integrate opinions from various departments in traditional consultations, and improves the completeness and usability of consultation results.

[0010] Optionally, in this embodiment, the consultation question is answered by each department agent, including: for each department agent, querying the professional knowledge block corresponding to the consultation question from the professional knowledge base configured for that department agent; using the professional knowledge block corresponding to the consultation question as the context of that department agent and inputting it into the large language model configured for that department agent; and generating the answer result of that department agent through the large language model of that department agent. In the implementation of the above solution, by configuring a dedicated professional knowledge base for each department agent and directly querying the professional knowledge block corresponding to the consultation question from that knowledge base, accurate filtering and targeted invocation of professional knowledge are achieved, avoiding the inefficiency of full search in a large-scale general knowledge base in traditional methods, significantly improving the accuracy and efficiency of knowledge retrieval, thereby enabling the rapid acquisition of the most relevant medical professional knowledge. In addition, by using the queried professional knowledge block as context input into the large language model, this solution creatively achieves a deep integration of professional knowledge and the capabilities of the large language model. This context injection approach not only ensures the professionalism and authority of the generated answers but also overcomes the inherent limitations of general-purpose large language models in the medical field, which may produce hallucinations or misinformation, significantly improving the reliability of the medical question-answering system. Furthermore, through a multi-agent system architecture where a central agent and agents from various departments work collaboratively, complex consultation tasks are decomposed, routed, and integrated. This allows the system to customize knowledge processing and answer generation based on the professional characteristics of different departments. This architecture fully leverages the domain expertise of each professional agent, ensuring the independence and accuracy of cross-disciplinary knowledge processing. Simultaneously, the central control agent enables process scheduling and result aggregation, enhancing the system's flexibility and scalability, and providing solid technical support for building an efficient and reliable multidisciplinary intelligent consultation system.

[0011] Optionally, in this embodiment of the application, the total agent summarizes the answers from each department agent, including: performing a consistency check on the answers from each department agent through the total agent to obtain a consistency check result; if there is a contradiction in the consistency check result, the total agent triggers interactive discussion among the department agents, and the final result of the interactive discussion and the answers from each department agent are summarized into a consultation result.

[0012] In implementing the above scheme, the consistency check of the responses from agents in each department by the central agent automatically identifies diagnostic discrepancies between different departments, avoiding potential omissions in manual verification and improving the accuracy of diagnostic results. Furthermore, by triggering interactive discussions between conflicting agents in different departments through the central agent, automated cross-departmental collaborative diagnosis is achieved, breaking down information silos between departments in traditional consultations. This allows for deep integration of medical opinions from different specialties, improving consultation efficiency. Moreover, by summarizing the final results of the interactive discussions and the responses from each department's agents into the consultation outcome, the consultation conclusion is ensured to retain both the professional opinions of each department and the consensus reached through discussion, making the final consultation result more comprehensive and authoritative.

[0013] This application also provides a device for predicting the risk of cardiac amyloidosis, comprising: a medical information acquisition module for acquiring the medical information of suspected patients; an information analysis and processing module for analyzing and processing the medical information using a large language model to obtain auxiliary prediction information including unsupported point information and triage suggestion results; a department doctor screening module for screening department doctors based on triage suggestion results; a feedback information receiving module for sending the auxiliary prediction information to the terminal device of the department doctor based on the triage suggestion results, so as to receive the doctor's feedback of new diagnostic information and subsequent medical information of suspected patients through the department doctor's terminal device, wherein the new diagnostic information includes the results of supplementary examination items based on unsupported point information; and a risk exclusion and confirmation module for analyzing and processing the subsequent medical information of suspected patients using a large language model if the new diagnostic information does not include the exclusion or confirmation of the risk of cardiac amyloidosis, until the unsupported point information in the obtained new diagnostic information no longer indicates the need for supplementary examination items; wherein the new diagnostic information and auxiliary prediction information are used to exclude or confirm the risk of cardiac amyloidosis.

[0014] Optionally, in this embodiment, the information analysis and processing module includes: a clinical indicator mapping submodule, used to extract key clinical indicators from medical information and map the key clinical indicators to standard terminology clinical indicators; a treatment knowledge acquisition submodule, used to perform an initial search in the treatment guideline knowledge base based on the standard terminology clinical indicators to obtain treatment guideline knowledge blocks; a path question generation submodule, used to generate a reasoning path and the precise retrieval question corresponding to the reasoning path through a large language model based on the treatment guideline knowledge blocks and medical information; a retrieval question conversion submodule, used to convert the precise retrieval question into a retrieval embedding vector; an embedding vector acquisition submodule, used to perform a second search in the vector database through the retrieval embedding vector to obtain an embedding vector knowledge block with a higher relevance to the medical information than the treatment guideline knowledge block, wherein the relevance between the embedding vector knowledge block and the medical information is greater than the relevance between the treatment guideline knowledge block and the medical information; and a prediction information generation submodule, used to input the embedding vector knowledge block and the medical information together into the large language model so that the large language model generates auxiliary prediction information.

[0015] Optionally, in this embodiment, the department doctor screening module includes: a doctor range determination submodule, used to determine the range of selectable doctors based on the department doctors' consultation time, shift schedule, holiday markers, and outpatient peak index; a doctor feature extraction submodule, used to extract selectable doctor features from the information of each selectable doctor within the selectable doctor range; a consultation success rate calculation module, used to calculate the consultation success rate of each selectable doctor based on the selectable doctor features and consultation medical information using an XGBoost model; and a department doctor screening submodule, used to screen department doctors from the selectable doctor range based on the consultation success rate.

[0016] Optionally, in this embodiment, the department doctor screening module further includes: a sample feature adding submodule, used to add pseudo-sample features with random rewards to each selectable doctor information in the range of selectable doctors; a sample feature merging submodule, used to merge the selectable doctor features and the pseudo-sample features with random rewards to obtain merged doctor features; and a doctor feature adding submodule, used to add the merged doctor features to the selectable doctor features and update the consultation success rate of each selectable doctor according to the selectable doctor features.

[0017] Optionally, in this embodiment of the application, the cardiac amyloidosis risk prediction device further includes: a consultation request receiving module, used to receive a consultation request sent by the terminal device of the department doctor; a consultation result obtaining module, used to initiate multi-Agent consultation according to the consultation request and obtain consultation results; and a consultation result sending module, used to send the consultation results to the terminal device of the department doctor.

[0018] Optionally, in this embodiment, the multiple agents include: a main agent and department agents; the consultation result acquisition module includes: a department agent determination submodule, used to parse the consultation question from the consultation request through the main agent, and determine the department agents that need to participate according to the consultation question; an answer result acquisition submodule, used to answer the consultation question through each department agent, and obtain the answer result of each department agent; and an answer result summarization submodule, used to summarize the answer results of each department agent through the main agent, and obtain the consultation result.

[0019] Optionally, in this embodiment, the answer result acquisition submodule includes: a professional knowledge query unit, used to query the professional knowledge block corresponding to the consultation question from the professional knowledge base configured by the department agent for each department agent; a professional knowledge input unit, used to input the professional knowledge block corresponding to the consultation question as the context of the department agent into the large language model equipped by the department agent; and an answer result generation unit, used to generate the answer result of the department agent through the large language model of the department agent.

[0020] Optionally, in this embodiment of the application, the answer result summarization submodule includes: an inspection result acquisition unit, used to perform a consistency check on the answer results of each department agent through the main agent to obtain a consistency check result; and a consultation result summarization unit, used to trigger interactive discussion among each department agent through the main agent if there is a contradiction in the consistency check results, and to summarize the final result of the interactive discussion and the answer results of each department agent into a consultation result.

[0021] This application also provides an electronic device, including a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the machine-readable instructions are executed by the processor to perform the methods described above.

[0022] This application also provides a computer-readable storage medium storing a computer program, which is executed by a processor to perform the methods described above.

[0023] This application also provides a computer program product, including: a computer program or computer instructions, which are executed by a processor to perform the method described above. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 The flowchart shown is a schematic diagram of the method for predicting the risk of cardiac amyloidosis provided in an embodiment of this application;

[0026] Figure 2 The diagram shown is a structural schematic of the cardiac amyloidosis risk prediction device provided in an embodiment of this application.

[0027] Figure 3 The diagram shows a structural schematic of an electronic device provided in an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in the embodiments of this application are for illustrative and descriptive purposes only and are not intended to limit the protection scope of the embodiments of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in the embodiments of this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of the embodiments of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0029] Furthermore, the described embodiments are merely a part of the embodiments of this application, and not all of them. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed embodiments of this application, but merely to illustrate selected embodiments of this application.

[0030] It is understood that the terms "first" and "second" in the embodiments of this application are used to distinguish similar objects. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" do not necessarily imply that they are different. In the description of the embodiments of this application, the term "and / or" is merely a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship. The term "multiple" refers to two or more (including two), and similarly, "multiple groups" refers to two or more groups (including two groups).

[0031] It should be noted that the cardiac amyloidosis risk prediction method provided in this application embodiment can be executed by an electronic device. Here, electronic device refers to a device terminal or server with the function of executing computer programs. Device terminals include, for example, smartphones, personal computers, tablets, personal digital assistants, or mobile internet devices. Servers refer to devices that provide computing services through a network. Servers include, for example, x86 servers and non-x86 servers. Non-x86 servers include, for example, mainframes, minicomputers, and UNIX servers.

[0032] Currently, the initial prediction process for common diseases is typically based on machine learning models trained on large-scale clinical data. These models can integrate relatively universal input data such as the patient's basic information, medical history, routine laboratory indicators, and imaging features, and output the disease risk probability through a single forward inference. Because these diseases, such as hypertension, diabetes, or coronary heart disease, have clear early biomarkers and high incidence rates, and their clinical characteristics are relatively typical, the models can run automatically in routine physical examinations or outpatient settings, achieving efficient and rapid initial screening without additional expert intervention. Thus, "one-time call, instant prediction" becomes possible.

[0033] However, cardiac amyloidosis is a rare disease with insidious clinical manifestations and multi-system involvement. Early symptoms are often confused with common heart diseases such as heart failure and hypertrophic cardiomyopathy. It lacks specific single diagnostic indicators, therefore, a universal model cannot currently be used to directly predict risk. Clinically, further evaluation is usually initiated only after a cardiologist suspects abnormal patient presentations (such as unexplained ventricular wall thickening or decreased cardiac function). This requires a multidisciplinary team consultation involving specialists in cardiology, hematology, and nephrology to comprehensively assess possible affected organs and sequentially perform targeted examinations such as serum free light chain testing, bone scans, cardiac MRI, or tissue biopsies to gradually rule out or confirm the diagnosis.

[0034] In practice, it has been found that the identification of cardiac amyloidosis relies on physicians' subjective suspicion and a linear screening process involving multidisciplinary collaboration, rather than an automated, standardized model-based initial screening mechanism. This process is time-consuming, resource-intensive, and highly dependent on expert experience, making it difficult to widely implement in asymptomatic or early stages. This passive and decentralized diagnostic model leads to a large number of potential cases being delayed or missed. Therefore, the overall efficiency of convening multiple outpatient medical teams for early risk prediction of rare diseases is significantly lower than that of automated model prediction processes for common diseases.

[0035] To improve the above issues, please see Figure 1 The illustrated flowchart represents a method for predicting the risk of cardiac amyloidosis provided in this application embodiment. The main idea of ​​this method is to automatically analyze patient medical information using a large language model, intelligently identify potential risk characteristics of cardiac amyloidosis, automatically generate auxiliary predictive information including supportive point information and triage suggestions, and push the screening results to the terminal devices of doctors in the corresponding departments. Then, based on newly added diagnostic information from doctors, the analysis is continuously optimized, forming a closed-loop feedback mechanism until the risk of cardiac amyloidosis is confirmed or ruled out. This effectively improves the problems of missed and misdiagnosed cases in the diagnosis of cardiac amyloidosis caused by the lack of specific symptoms, the need for multidisciplinary collaboration, and the long diagnostic cycle, significantly improving the efficiency of early risk prediction for rare diseases. The implementation methods of the above-mentioned method for predicting the risk of cardiac amyloidosis may include:

[0036] Step S110: Obtain the medical information of suspected patients.

[0037] Suspected cases refer to patients who may have cardiac amyloidosis or other related diseases, but have not yet been diagnosed and need further analysis and confirmation through medical information. Examples include patients with cardiac discomfort symptoms but no clear cause, patients whose physical examination reveals abnormal cardiac indicators, or high-risk individuals with a family history of cardiac amyloidosis.

[0038] Medical information refers to all relevant medical data generated when a suspected patient seeks medical treatment at a medical institution. This data is used to assist in diagnosis and analysis and may include medical information generated during the current visit of the suspected patient or information generated in the past.

[0039] Step S120: Analyze and process the medical information of patients through a large language model to obtain auxiliary prediction information that includes information on unsupported points and triage suggestions.

[0040] Large Language Models (LLMs) are artificial intelligence-based natural language processing models used to analyze and process medical information and generate auxiliary predictive information. Examples of LLMs include GPT-4, BERT, PaLM, or customized language models specifically trained for the medical field—large pre-trained language models. It's understandable that by analyzing and processing medical information through large language models, they can automatically identify discrepancies and provide triage suggestions, significantly improving diagnostic efficiency and accuracy. The powerful natural language processing capabilities of large language models can quickly parse complex medical information, reducing the time and errors of manual analysis and ensuring the scientific rigor and consistency of the diagnostic process.

[0041] Unsupported information refers to missing or insufficient information identified by the large language model after analyzing the medical information received during the consultation, which is insufficient to support the diagnosis or exclusion of cardiac amyloidosis. Examples of unsupported information include: missing results of specific biomarker tests, insufficient resolution of imaging examinations, or incomplete medical history records. This unsupported information can be generated through logical analysis of the medical information received during the consultation and comparison with medical guidelines by the large language model.

[0042] Auxiliary predictive information refers to the comprehensive analysis results output by the large language model, which may include: disease risk scores, supporting point information, unsupporting point information, and triage suggestions. It is used to assist doctors in decision-making and may include text reports, structured data tables, or visualization charts.

[0043] Step S130: Select departmental doctors based on the triage recommendations.

[0044] Triage recommendations refer to suggestions generated by a large language model based on medical information regarding which department a suspected patient should consult for further treatment. For example, a recommendation might suggest referral to cardiology, neurology, or genetic counseling. Understandably, by selecting doctors based on these recommendations and sending supplementary predictive information to doctors' devices, precise matching and efficient utilization of medical resources are achieved. This intelligent triage and recommendation mechanism avoids the subjectivity and uncertainty of traditional triage processes, ensuring that patients receive the most appropriate medical resources and services.

[0045] The department doctors are selected based on the triage recommendations and whose professional fields are related to the suspected patient's condition, such as cardiologists, neurologists, and radiologists.

[0046] Step S140: Based on the triage recommendation results, the auxiliary prediction information is sent to the terminal device of the department doctor so that the department doctor can receive the new diagnostic information and the follow-up medical information of suspected patients through the terminal device. The new diagnostic information includes the results of the examination items supplemented based on the information of unsupported points.

[0047] Terminal devices are electronic devices used by doctors in a department to receive and view auxiliary predictive information and input feedback information, such as hospital workstation computers, tablets, or smartphones.

[0048] New diagnostic information refers to new medical data obtained by doctors in the department after supplementing examinations based on the unsupported information in the auxiliary predictive information. Examples include new biopsy results, supplementary laboratory reports (such as blood tests and urine tests), and imaging examination results (such as echocardiography and MRI). This new diagnostic information can be generated by the laboratory or radiology department by doctors ordering new examinations.

[0049] Subsequent medical information refers to all new medical-related information generated after the initial visit of a suspected patient, such as follow-up test results, hospitalization observation records, or medication reaction records. This subsequent medical information is obtained through continuous medical monitoring and examination.

[0050] The results of the examination are output data for specific medical examinations that do not support supplemental point information, such as serum free light chain test values, cardiac amyloidosis PET-CT images, and tissue biopsy pathology reports. These results can be obtained through laboratory tests, imaging examinations, or pathological examinations.

[0051] Step S150: If the newly added diagnostic information does not include the risk of excluding or confirming cardiac amyloidosis, the subsequent medical information of the suspected patient is analyzed and processed through the large language model until the unsupported information in the newly obtained diagnostic information no longer indicates the need for supplementary examinations; among them, the newly added diagnostic information and auxiliary predictive information are used to exclude or confirm the risk of cardiac amyloidosis.

[0052] Cardiac amyloidosis risk is an assessment of the probability or likelihood that a suspected patient has cardiac amyloidosis, such as high risk, medium risk, or low risk.

[0053] Understandably, if the unsupported information in the newly added diagnostic information still indicates the need for additional examinations, then the departmental physician can supplement the examinations based on the unsupported information in the auxiliary predictive information to obtain new medical data, such as newly performed biopsy results, supplementary laboratory reports (e.g., blood tests, urine tests), and imaging examination results (e.g., echocardiography and MRI). This new diagnostic information can then be sent to the electronic device via the physician's terminal. The combined use of the newly added diagnostic information and auxiliary predictive information provides double assurance for ruling out or confirming the risk of cardiac amyloidosis. This synergistic effect not only improves diagnostic accuracy but also reduces the possibility of missed diagnoses and misdiagnoses, thereby improving overall medical quality and patient safety.

[0054] In implementing the above solution, a large language model is used to automatically analyze patient medical information, intelligently identify potential risk characteristics of cardiac amyloidosis, and automatically generate auxiliary predictive information including unsupported information and triage suggestions, rather than relying on the doctor's subjective suspicion. This intelligently recommends relevant departments (such as cardiology, hematology, and nephrology) for consultation and pushes the screening results to the corresponding doctors' terminals. The analysis is then continuously optimized based on the doctor's added diagnostic information, forming a closed-loop feedback mechanism until the risk of cardiac amyloidosis is confirmed or ruled out. This closed-loop feedback mechanism continuously corrects and improves diagnostic information, ensuring optimal diagnostic results when unsupported information no longer prompts for additional examinations. This significantly improves the comprehensiveness and reliability of the diagnosis, realizing a shift from "doctor's proactive suspicion" to "system-driven screening." It effectively addresses the problems of missed and misdiagnosed cases in the diagnosis of cardiac amyloidosis caused by the lack of specific symptoms, the need for multidisciplinary collaboration, and long diagnostic cycles, significantly improving the efficiency of early risk prediction for rare diseases.

[0055] Optionally, as an alternative implementation of step S110 above, for example, the hospital's electronic equipment (e.g., a server) can obtain the suspected patient's medical information from the Electronic Medical Record (EMR), Hospital Information System (HIS), or Laboratory Information System (LIS). Specifically, the aforementioned medical information may include medical record information, examination information, laboratory reports (such as blood tests and urine tests), and imaging examination results (such as echocardiography and MRI). For example, the medical record information might include: a patient admitted to the hospital due to "weakness in the limbs for 2 years, worsening with numbness and pain for 1+ years." Outpatient examination results: [Free light chain κ type / Free light chain λ type 0.7089. ][2023-08-15][Immunofixation electrophoresis][2023-08-15] IgA κ type M protein negative (-). Echocardiography: [(Emergency) B-type natriuretic peptide (BNP)][2023-08-14] B-type natriuretic peptide 41.4 pg / mL. Left and right ventricular hypertrophy. Strain imaging results: Overall LV-GLS: -18.3%; basal segment approximately -15.7% (lowered ventricular septum base), mid-segment approximately -17%, apical segment approximately -22%; mild tricuspid regurgitation. Atrial septal thickening. Uneven thickening of the left ventricular septal wall and inferior wall, with the thickest point located in the ventricular septum, diastolic thickness approximately 1.7 cm. No obvious abnormalities were seen in first-pass perfusion. Patchy abnormal enhancement was seen in the septal wall and inferior wall of the left ventricle on delayed contrast-enhanced CT. Sinus rhythm, left axis deviation (-68°). Left anterior fascicular block.

[0056] As an optional implementation of step S120 above, the above implementation of analyzing and processing medical information through a large language model may include:

[0057] Step S121: Extract key clinical indicators from medical information and map them to standard terminology clinical indicators.

[0058] Key clinical indicators are medical indicators or features that are of great significance for diagnosis and treatment extracted from medical information during a patient visit, such as blood glucose levels, white blood cell counts, blood pressure values, and pain scores. These key clinical indicators can be extracted from unstructured medical texts using natural language processing technology or obtained directly from structured medical databases.

[0059] The implementation of step S121 above is as follows: Assume the patient's medical record information is as follows: The patient was admitted to the hospital due to "weakness in the limbs for 2 years, worsening with numbness and pain for 1+ years". Outpatient examinations: [Free light chain κ type / Free light chain λ type 0.7089. ][2023-08-15][Immunofixation electrophoresis][2023-08-15] IgA κ type M protein negative (-). Echocardiography: [(Emergency) B-type natriuretic peptide (BNP)][2023-08-14] B-type natriuretic peptide 41.4 pg / mL. Left and right ventricular hypertrophy. Strain imaging results: Overall LV-GLS: -18.3%; basal segment approximately -15.7% (lowered ventricular septum base), mid-segment approximately -17%, apical segment approximately -22%; mild tricuspid regurgitation. Atrial septal thickening. The left ventricular septum and inferior wall show uneven thickening, with the thickest point located in the ventricular septum. The diastolic thickness is approximately 1.7 cm. First-pass perfusion showed no significant abnormalities. Delayed contrast-enhanced CT scans revealed patchy abnormal enhancement of the septal wall and the inferior wall of the left ventricle. The patient has sinus rhythm with left axis deviation (-68°). Left anterior fascicular block is present. Key clinical indicators (CRIs) can be extracted from the patient's medical history using Named Entity Recognition (NER) models such as BiLSTM-CRF, RoBERTa, or BERT-CRF. These CRIs can include key laboratory indicators, laboratory test results, and clinical clues. Then, these CRIs can be mapped to standard terminology for clinical indicators. Among them, the standard clinical indicators here include: ["Free light chain κ type / free light chain λ type: 0.7089", "IgAκM protein negative", "BNP 41.4pg / mL", "Left ventricular hypertrophy", "Atrial septal thickening", "Left ventricular septal wall thickness 1.7cm", "Delayed enhancement abnormal enhancement", "Overall LV-GLS -18.3%", "Myocardial strain basal segment damage", "Mild tricuspid regurgitation", "Left anterior fascicular block", "Myocardial infiltrative lesions", "Suspected cardiac amyloidosis"].

[0060] Understandably, extracting key clinical indicators from medical records and mapping them to standardized terminology can eliminate ambiguity and non-standardized expressions in medical information, thereby improving the accuracy of subsequent retrieval and analysis. This step ensures data consistency and comparability, providing high-quality input for subsequent clinical guideline retrieval and large language model processing.

[0061] Step S122: Perform an initial search in the clinical indicator knowledge base based on standard terminology to obtain clinical indicator knowledge blocks.

[0062] Standardized clinical terminology is a standardized representation that maps key clinical indicators to a unified medical terminology system, facilitating subsequent knowledge retrieval and processing.

[0063] A clinical practice guideline knowledge base refers to a knowledge base that stores standardized clinical practice guidelines, including diagnostic and treatment recommendations for various diseases, such as guidelines for rare diseases or cardiac amyloidosis. This knowledge base can be obtained by collecting and organizing clinical practice guidelines from authoritative medical institutions and storing them in a structured, searchable knowledge block format.

[0064] A treatment guideline knowledge block is a specific knowledge unit in the treatment guideline knowledge base, containing diagnostic and treatment recommendations for specific diseases or symptoms. Specifically, it can be a knowledge block that breaks down treatment guidelines into independent vector knowledge units according to topics or disease classifications.

[0065] For example, the implementation of step S122 above can be as follows: keywords can be decomposed based on standard terminology and clinical indicators to obtain decomposed clinical indicator keywords. Then, the decomposed clinical indicator keywords can be searched for the first time in the diagnosis and treatment guideline knowledge base to obtain the diagnosis and treatment guideline knowledge block. The specific content of the diagnostic and treatment guidelines knowledge block may include: [{"doc_id": "doc_cardiac_AL_amyloidosis","section": "§2.1","span_text": "Cardioamyloidosis often presents as ventricular wall thickening, atrial septal thickening, and a significant decrease in myocardial strain at the basal segment while the apical segment is preserved (apical sparing sign)."},{"doc_id": "doc_FLC_reference","section": "§1.3","span_text": "Abnormal free light chain κ / λ ratio is a key clue for the diagnosis of AL amyloidosis. If immunofixation electrophoresis is negative but the κ / λ ratio is abnormal, it should be highly suspected."},{"doc_id": "doc_cardiac_MRI_guidelines","section": "§3.2","span_text": "Delayed enhancement on cardiac MRI suggests myocardial interstitial infiltration; amyloidosis commonly shows patchy or diffuse enhancement."}]. Understandably, conducting an initial search in the clinical guidelines knowledge base based on standard terminology can quickly locate the knowledge blocks in the guidelines most relevant to the patient's condition, providing a preliminary reference framework for subsequent precise searches and reasoning, thereby improving search efficiency and avoiding blind searches in a large-scale knowledge base.

[0066] Step S123: Generate a reasoning path and the corresponding precise retrieval question based on the knowledge blocks of the diagnosis and treatment guidelines and the medical information of the patient through the large language model.

[0067] The reasoning path is the logical reasoning process generated by the large language model based on medical information and treatment guidelines. It is used to explain how to derive a diagnosis or treatment suggestion from the input information. This reasoning path can be generated by the reasoning ability of the large language model and is usually presented in the form of natural language or logical chains.

[0068] The precise retrieval question is a targeted question generated by the large language model based on the reasoning path, used to further retrieve more relevant medical knowledge. This precise retrieval question can be automatically generated by the large language model based on the information needs of the current reasoning stage.

[0069] The implementation of steps S123 above can be exemplified as follows: The large language model in the Agent can call the Reasoning Path Tool and Question Generator to generate a reasoning path based on the knowledge blocks of the treatment guidelines and the patient's medical information. Then, based on this reasoning path, a precise retrieval question corresponding to the desired result can be generated. Examples of reasoning paths include: 1. BNP: 41.4 pg / mL (mildly elevated / near normal), suggesting a possible early stage of cardiac involvement. 2. Echocardiography: Left and right ventricular hypertrophy + atrial septal thickening + LV-GLS -18.3% (basal segment reduction, apical segment preservation) → a typical apical sparing sign. 3. MRI: Delayed patchy enhancement of the septal and inferior walls → supporting infiltrative lesions, commonly seen in cardiac amyloidosis. 4. Neurological manifestations: Long-term limb weakness + numbness and pain → suggesting possible peripheral nerve involvement, consistent with one of the systemic involvement characteristics of amyloidosis. 5. Immunological indicators: FLCκ / λ ratio 0.71, within the reference range of 0.26–1.65 → superficially normal, but needs to be considered in conjunction with clinical findings; negative immunofixation electrophoresis → highly likely to rule out AL. Overall: Clinical symptoms (neurological), imaging (cardiac), immunology (suspected light chain) → possibly ATTR-type cardiac amyloidosis.

[0070] Among them, the precise retrieval question corresponding to the above-mentioned generative reasoning path is, for example: 1. Typical ultrasound signs of cardiac amyloidosis ( apical sparing 1. What are the sensitivity and specificity of BNP? 2. If immunofixation electrophoresis is negative but the κ / λ ratio is on the edge of the normal range, is it still possible to have AL type amyloidosis? 3. What is the diagnostic value and staging significance of BNP in early cardiac amyloidosis? 4. Decision: Do the questions arising after reasoning require further investigation?

[0071] Step S124: Convert the exact retrieval question into a retrieval embedding vector.

[0072] Retrieving embedding vectors transforms the precise retrieval question into a high-dimensional vector representation for semantic similarity retrieval in a vector database. For example, embedding models (such as BERT and Sentence-BERT) can convert textual questions into vector representations. In this process, by converting the precise retrieval question into retrieval embedding vectors and then performing a second retrieval in the vector database, vector similarity calculations can be used to find embedding vector knowledge blocks with higher relevance to the medical information. This step, through vectorized retrieval technology, achieves deep matching of unstructured medical knowledge, improving the accuracy and relevance of retrieval results.

[0073] Step S125: By retrieving the embedded vector, perform another search in the vector database to obtain the embedded vector knowledge block that has a higher relevance to the medical information and the treatment guideline knowledge block. The relevance between the embedded vector knowledge block and the medical information is greater than that between the treatment guideline knowledge block and the medical information.

[0074] Vector databases are databases that store vector representations of knowledge blocks and support efficient retrieval based on vector similarity. Examples of vector databases include Pinecone, Milvus, and FAISS. They can convert knowledge block text into vectors through an embedding model and then store them in a dedicated vector database.

[0075] Embedded vector knowledge blocks are knowledge units that are highly relevant to medical information retrieved from a vector database and stored in vector form. These embedded vector knowledge blocks can be obtained by converting medical knowledge text into vectors and storing them in the vector database. They can also be used for similarity retrieval.

[0076] For example, if the content output by the large language model in the Agent indicates that a further retrieval is still needed in the vector database, the Agent can transform the precise retrieval problem into retrieving the embedding vector. By retrieving the embedding vector, a further retrieval is performed in the vector database to obtain an embedding vector knowledge block that has a higher relevance to the medical information than the medical guideline knowledge block. The relevance between the embedding vector knowledge block and the medical information is greater than the relevance between the medical guideline knowledge block and the medical information.

[0077] Step S126: Input the embedded vector knowledge blocks and medical information from the visit into the large language model so that the large language model can generate auxiliary prediction information.

[0078] The auxiliary prediction information is the medical advice ultimately generated by the large language model, including possible diagnoses and treatment plans. It can be natural language text generated by the large language model based on embedded vector knowledge blocks and medical information.

[0079] One implementation of step S126 above is as follows: First, the medical information is converted into an embedded medical information vector. Then, an Agent inputs the embedded vector knowledge block and the embedded medical information vector into a large language model, enabling the large language model to generate auxiliary prediction information. In this implementation, by inputting the embedded vector knowledge block and the medical information into the large language model to generate auxiliary prediction information, personalized treatment suggestions can be generated by combining the latest medical knowledge with the patient's specific situation. This step fully utilizes the comprehensive reasoning ability of the large language model, making the generated auxiliary prediction information more accurate and reliable.

[0080] As an optional implementation of step S130 above, the implementation of screening department doctors based on triage recommendation results may include:

[0081] Step S131: Determine the range of available doctors based on the department's doctors' consultation time, shift schedule, holiday markings, and outpatient peak index.

[0082] The range of selectable doctors is a set of doctors that meet the basic criteria initially selected based on the triage recommendations. This range of selectable doctors can be filtered from the doctor database based on factors such as the doctor's schedule and specialty.

[0083] The implementation of step S131 above is as follows: The SQL query template is populated with the doctors' consultation times, shift schedules, holiday markers, and peak outpatient index to obtain the SQL query statement. Then, the SQL query statement is executed in the relational database to obtain the range of selectable doctors. It is understandable that determining the range of selectable doctors by comprehensively considering multiple dimensions such as consultation times, shift schedules, holiday markers, and peak outpatient index can more comprehensively reflect the actual capacity of doctors to treat patients, avoiding screening bias caused by relying solely on a single factor such as doctor's professional title or experience, and improving the accuracy and rationality of doctor selection.

[0084] Step S132: Extract the features of the available doctors from the information of each available doctor within the range of available doctors.

[0085] Selectable doctor features are features extracted from available doctor information, such as structured features extracted from a doctor information database. These structured features can be used to assess a doctor's suitability for treatment, such as the doctor's title, specialization, number of patients seen, and patient satisfaction rating.

[0086] Step S133: Use the XGBoost model to calculate the consultation success rate of each available doctor based on the characteristics of the available doctors and the medical information of the patients.

[0087] The XGBoost model is a machine learning model that predicts the match between patients and doctors. XGBoost itself is a gradient boosting decision tree algorithm used to calculate the success rate of a consultation based on doctor characteristics and medical information.

[0088] The implementation of step S133 above is as follows: First, the XGBoost model is trained using all historical triage logs. These historical triage logs may include: historical patient information (such as patient age, gender, disease warning score, doctor preferences, etc.) and doctor information (doctor's seniority, historical queue length, number of similar cases treated, prognosis and satisfaction scores of historical patients, etc.). This allows the trained XGBoost model to calculate the consultation success rate of each available doctor based on the features of the available doctors and the medical information of the patient's visit. The consultation success rate is the probability that a doctor can successfully resolve the current patient's medical problem, which can be calculated by a machine learning model (such as XGBoost) based on doctor features and patient information. By extracting features from the available doctor information and combining them with the XGBoost model to calculate the consultation success rate, the powerful feature processing and prediction capabilities of machine learning algorithms are utilized. This allows for the uncovering of the complex nonlinear relationship between doctor features and consultation success rate, significantly improving the accuracy of doctor recommendations compared to traditional manual experience judgment or simple rule matching.

[0089] Step S134: Select doctors from the available doctors based on the success rate of the consultation.

[0090] For example, step S134 can be implemented by sorting the range of available doctors according to their success rates from highest to lowest, and then selecting the doctor with the highest success rate from the range of available doctors; or, randomly selecting a doctor from multiple doctors whose success rates are greater than a preset success rate threshold. It is understood that by filtering based on the calculated success rate, the matching degree of doctors is quantified and ranked, ensuring that the final doctor recommendation result has objective data support, reducing subjective interference, and simultaneously improving the success rate and satisfaction of patient visits.

[0091] As an optional implementation of step S132 above, after extracting the selectable doctor features from the selectable doctor information within the selectable doctor range, the method may further include:

[0092] Step S132a: Add pseudo-sample features with random rewards to the information of each selectable doctor in the range of selectable doctors.

[0093] Pseudo-sample features are doctor features added artificially with random rewards. Specifically, a random number generator can be used to add a small amount of random numerical features to each doctor.

[0094] An example implementation of step S132a above is as follows: The Follow The Perturbed Leader (FTPL) framework is used to add pseudo-sample features with random rewards to the information of each available doctor within the range of selectable doctors. This can be understood as temporarily adding a few pseudo-sample features with random rewards to each doctor using the FTPL framework, essentially rolling a die on the existing data, giving doctors with fewer samples and uncertain information additional volatility. It is understood that adding pseudo-sample features with random rewards to the information of each selectable doctor can effectively alleviate the cold start problem in the recommender system, because the introduction of random rewards can increase the diversity of the data, allowing the model to access a wider range of feature distributions in the early stages of training, thereby converging to a better initial state more quickly.

[0095] Step S132b: Merge the optional doctor features and the pseudo-sample features of the random reward to obtain the merged doctor features.

[0096] The merged doctor features are a new feature vector resulting from the combination of the original optional doctor features and the pseudo-sample features. Understandably, merging the optional doctor features and the pseudo-sample features with random rewards enhances the model's robustness because the addition of random rewards simulates uncertainty in real-world scenarios, allowing the model to maintain good performance even when faced with noisy data or anomalies.

[0097] Step S132c: Add the merged doctor features to the selectable doctor features, and update the consultation success rate of each selectable doctor based on the selectable doctor features.

[0098] The implementation of steps S132b to S132c above is as follows: The selectable doctor features and the pseudo-sample features of the random reward are merged to obtain merged doctor features. An incremental least squares or logistic regression is then performed on the merged doctor features to obtain updated doctor features. These updated doctor features are then added to the selectable doctor features, and the consultation success rate of each selectable doctor is updated based on these features. In the implementation of this scheme, by adding the merged doctor features to the selectable doctor features and updating the consultation success rate, the recommendation weight of doctors can be dynamically optimized. This is because the introduction of random rewards can balance the relationship between exploration and utilization, allowing the system to recommend doctors with high success rates while also exploring potential high-value doctors.

[0099] As an optional implementation of step S130 above, after screening departmental doctors based on triage recommendations, the following may also be included:

[0100] Step S135: Receive consultation requests sent from the terminal devices of departmental doctors.

[0101] A consultation request is a multi-departmental collaborative diagnosis and treatment request initiated by a physician in a department through a terminal device. Physicians can fill out a consultation application form and submit it through the workstation system.

[0102] For example, in implementing step S135 above: if a physician in a department believes their existing knowledge is insufficient and requires supplementation from physicians in other departments, they can initiate a multi-departmental collaborative consultation request through their terminal device. The terminal device will send this consultation request to an electronic device (such as a server). The electronic device (such as a server) can receive the consultation request sent by the physician's terminal device. It is understood that by receiving the consultation request sent by the physician's terminal device and initiating multi-agent consultation, cross-departmental and cross-professional collaborative diagnosis can be achieved, thereby improving the accuracy and comprehensiveness of diagnosis. Especially in the diagnosis of complex cases, it can integrate knowledge and experience from multiple professional fields, reducing the limitations of a single doctor's subjective judgment.

[0103] Step S136: Initiate multi-agent consultation based on the consultation request and obtain consultation results.

[0104] A multi-agent system is a consultation system composed of multiple intelligent agents, each representing knowledge in a specific medical specialty. Understandably, the introduction of this multi-agent system makes the consultation process more intelligent and automated, enabling diagnostic reasoning based on pre-defined rules and algorithms, reducing human error and omissions. Simultaneously, it allows for continuous optimization of the diagnostic model based on historical data and cases, improving the system's adaptability and diagnostic accuracy.

[0105] Step S137: Send the consultation results to the terminal device of the department doctor.

[0106] Understandably, by sending consultation results to the terminal devices of doctors in the department, real-time feedback and sharing of diagnostic results are achieved, ensuring that doctors can obtain consultation opinions in a timely manner, thereby accelerating the formulation and implementation of treatment decisions, improving the response speed of medical services and patient satisfaction.

[0107] As an optional implementation of step S132 above, the aforementioned multi-Agent includes: a main agent and agents for each department; the implementation of activating multiple agents for consultation based on a consultation request may include:

[0108] Step S136a: Parse the consultation question from the consultation request through the main agent, and determine the department agents that need to participate based on the consultation question.

[0109] The implementation of step S136a above is as follows: A central agent parses the consultation question from the consultation request, and determines and coordinates the relevant departmental agents based on the consultation question. This central agent is the coordinator in the multi-agent system and the main control module of the consultation system, primarily responsible for parsing the consultation question, allocating tasks, and summarizing results. It can be understood that by setting up a hierarchical structure of the central agent and departmental agents, intelligent allocation and collaborative processing of consultation questions are achieved. The central agent can automatically parse the consultation request and determine the relevant departmental agents, avoiding the inefficiency and potential omissions of manual allocation, and improving the accuracy and comprehensiveness of the consultation.

[0110] Step S136b: Obtain the response results from each department's agent by having the agent answer the consultation questions.

[0111] Each department agent is an intelligent entity representing a specific medical specialty, responsible for answering consultation questions in its own field, such as a cardiology agent, a hematology agent, and a nephrology agent.

[0112] For example, in implementing step S136b above, each department agent can use Backward Retrieval-Augmented Generation (BRAG) technology to answer consultation questions and obtain the answer results of each department agent, thereby reducing illusions in the reasoning process of the large language model set in each department agent. Specifically, in the process of answering consultation questions, a department agent can first search the vector database of treatment guideline knowledge blocks such as Pinecone, Milvus, and FAISS based on the keywords of the consultation question and the medical information, and obtain the search results. Then, based on the search results and the medical information, it constructs and generates a reasoning path and the precise search question corresponding to the reasoning path, and converts the precise search question into a search embedding vector. Finally, it searches the vector database again through the search embedding vector to obtain the embedding vector knowledge block with a higher relevance to the medical information than the treatment guideline knowledge block. The embedding vector knowledge block and the medical information are then input together into the large language model of the department agent, so that the large language model of the department agent can generate auxiliary prediction information.

[0113] Understandably, the departmental agents possess the ability to utilize various tools, including Extractor (information extraction), Retriever (medical knowledge retrieval), ReasoningPath Tool (reasoning path generation), Question Generator (question generation), and Summarizer (diagnostic summary tool), for different consultation questions. Furthermore, by having each departmental agent independently process consultation questions and generate responses, this parallel processing approach significantly improves consultation efficiency and shortens overall response time. Simultaneously, each departmental agent can focus on answering questions within its own area of ​​expertise, ensuring the professionalism and reliability of its responses.

[0114] As another optional implementation of step S136b above, the implementation method of answering consultation questions through department agents may include: for each department agent, querying the professional knowledge block corresponding to the consultation question from the professional knowledge base configured for that department agent. It is understood that by configuring a professional knowledge base for the department agents and directly querying the professional knowledge block corresponding to the consultation question from that knowledge base, precise filtering and targeted invocation of professional knowledge are achieved. This avoids the inefficiency of performing a full search in a large-scale general knowledge base in traditional methods, significantly improving the accuracy and efficiency of knowledge retrieval, thereby enabling the rapid acquisition of the most relevant medical professional knowledge.

[0115] Then, the professional knowledge block corresponding to the consultation question is used as the context of the department's agent and input into the large language model configured for that department's agent. Understandably, by inputting the retrieved professional knowledge block as context into the large language model, this solution creatively achieves a deep integration of professional knowledge and the capabilities of the large language model. This context injection method not only ensures the professionalism and authority of the generated answer but also overcomes the inherent defects of general-purpose large language models in the medical field, which may produce illusions or misinformation, significantly improving the reliability of the medical question-answering system.

[0116] Finally, the system generates the department agent's response using a large language model. The professional knowledge base is a set of specialized medical knowledge configured for a specific department agent. This base can be constructed by staff through collecting and organizing authoritative departmental guidelines and literature, creating a structured or vectorized knowledge base for the corresponding department agent to access. Professional knowledge blocks are specific knowledge units within the professional knowledge base, used to answer specific professional questions. These blocks can be created by breaking down professional knowledge into independent knowledge units by topic. In implementing this solution, the architecture design of configuring a large language model for the department agent enables the system to customize knowledge processing and response generation based on the professional characteristics of different departments. This modular design not only improves the system's flexibility and scalability but also achieves effective isolation and specialized processing of cross-departmental knowledge, providing a technical foundation for building a multidisciplinary collaborative intelligent consultation system.

[0117] Step S136c: The overall agent summarizes the responses from agents in each department to obtain the consultation results.

[0118] As an optional implementation of step S136c above, the implementation of summarizing the response results of each department's agents through the overall agent may include: performing a consistency check on the response results of each department's agents through the overall agent to obtain a consistency check result. It is understood that performing a consistency check on the response results of each department's agents through the overall agent can automatically identify diagnostic inconsistencies between different departments, avoiding potential omissions in manual verification and improving the accuracy of diagnostic results.

[0119] If the consistency check results are contradictory, the main agent triggers interactive discussions among the departmental agents. The final results of these discussions, along with the responses from each departmental agent, are then combined to form the consultation outcome. The consistency check is the process by which the main agent logically verifies the consistency of the responses from each departmental agent. This can be achieved through a rule engine or the logical reasoning capabilities of a large language model. The interactive discussion is a multi-agent negotiation process coordinated by the main agent when contradictory responses exist. In essence, by triggering interactive discussions among contradictory departmental agents through the main agent, automated cross-departmental collaborative diagnosis is achieved. This breaks down information silos between departments in traditional consultations, allowing for deep integration of medical opinions from different specialties and improving consultation efficiency.

[0120] Please see Figure 2 The diagram shown is a structural schematic of the cardiac amyloidosis risk prediction device provided in an embodiment of this application; this embodiment of the application provides a cardiac amyloidosis risk prediction device 200, including:

[0121] The medical information acquisition module 210 is used to acquire the medical information of suspected patients.

[0122] The information analysis and processing module 220 is used to analyze and process medical information from patients through a large language model to obtain auxiliary predictive information that includes information on unsupported points and triage suggestions.

[0123] The Department Doctor Screening Module 230 is used to screen department doctors based on triage recommendations.

[0124] The feedback information receiving module 240 is used to send auxiliary prediction information to the terminal device of the department doctor based on the triage suggestion result, so that the department doctor can receive the new diagnostic information and the follow-up medical information of suspected patients through the terminal device of the doctor. The new diagnostic information includes the results of examination items supplemented based on the information of unsupported points.

[0125] The risk exclusion and diagnosis module 250 is used to analyze and process the subsequent medical information of suspected patients through a large language model if the newly added diagnostic information does not include the risk of excluding or confirming cardiac amyloidosis, until the unsupported points in the newly obtained diagnostic information no longer indicate the need for supplementary examinations; among them, the newly added diagnostic information and auxiliary prediction information are used to exclude or confirm the risk of cardiac amyloidosis.

[0126] As an optional implementation of the above-mentioned device, the information analysis and processing module includes:

[0127] The Clinical Indicator Mapping Submodule is used to extract key clinical indicators from medical information and map these key clinical indicators to standard terminology clinical indicators.

[0128] The clinical knowledge acquisition submodule is used to perform an initial search in the clinical guidelines knowledge base based on standard terminology and clinical indicators to obtain clinical guidelines knowledge blocks.

[0129] The path question generation submodule is used to generate inference paths and corresponding precise retrieval questions based on the knowledge blocks of diagnosis and treatment guidelines and medical information from a large language model.

[0130] The retrieval question transformation submodule is used to convert the exact retrieval question into a retrieval embedding vector.

[0131] The embedding vector acquisition submodule is used to retrieve embedding vectors and then search the vector database again to obtain embedding vector knowledge blocks that have a higher relevance to the medical information than the treatment guideline knowledge blocks. The relevance between the embedding vector knowledge blocks and the medical information is greater than that between the treatment guideline knowledge blocks and the medical information.

[0132] The prediction information generation submodule is used to input the embedded vector knowledge blocks and medical information from the patient visit into the large language model, so that the large language model can generate auxiliary prediction information.

[0133] As an optional implementation of the above-mentioned device, the departmental doctor screening module includes:

[0134] The Doctor Scope Determination submodule is used to determine the range of available doctors based on the department's doctors' consultation time, shift schedule, holiday markings, and outpatient peak index.

[0135] The doctor feature extraction submodule is used to extract the features of each available doctor from the information of each available doctor within the range of available doctors.

[0136] The consultation success rate calculation module is used to calculate the consultation success rate of each available doctor based on the characteristics of available doctors and medical information using the XGBoost model.

[0137] The Department Doctor Screening Submodule is used to select department doctors from the available doctors based on the success rate of consultations.

[0138] As an optional implementation of the above-mentioned device, the departmental doctor screening module further includes:

[0139] The Sample Feature Addition submodule is used to add pseudo-sample features with random rewards to the information of each selectable doctor in the range of selectable doctors.

[0140] The sample feature merging submodule is used to merge the optional doctor features and the pseudo-sample features of the random reward to obtain the merged doctor features.

[0141] The Doctor Feature Addition submodule is used to add the merged doctor features to the selectable doctor features and update the consultation success rate of each selectable doctor based on the selectable doctor features.

[0142] As an optional embodiment of the above-mentioned device, the cardiac amyloidosis risk prediction device further includes:

[0143] The consultation request receiving module is used to receive consultation requests sent by the terminal devices of doctors in the department.

[0144] The consultation result acquisition module is used to initiate multi-agent consultations based on consultation requests and obtain consultation results.

[0145] The consultation result sending module is used to send consultation results to the terminal devices of doctors in the department.

[0146] As an optional implementation of the above-mentioned device, the multi-agent includes: a main agent and agents for each department; the consultation result acquisition module includes:

[0147] The Department Agent Determination Submodule is used to parse the consultation question from the consultation request through the main Agent, and determine the department agents that need to participate based on the consultation question.

[0148] The response result acquisition submodule is used to obtain the response results from each department's agent by having the agent answer the consultation questions.

[0149] The response result summary submodule is used to summarize the response results of agents from various departments through the main agent to obtain the consultation results.

[0150] As an optional implementation of the above-mentioned device, the result acquisition submodule includes:

[0151] The professional knowledge query unit is used to query the professional knowledge block corresponding to the consultation question from the professional knowledge base configured for each department agent.

[0152] The professional knowledge input unit is used to input the professional knowledge block corresponding to the consultation question as the context of the department agent into the large language model configured for the department agent.

[0153] The answer generation unit is used to generate the answer results of the department agent through the large language model of the department agent.

[0154] As an optional implementation of the above-mentioned device, the result summarization submodule includes:

[0155] The inspection result acquisition unit is used to perform consistency checks on the response results of agents in each department through the overall agent, and obtain consistency check results.

[0156] The consultation result summary unit is used to trigger interactive discussions among the agents of each department through the main agent if there are contradictions in the consistency check results, and to summarize the final results of the interactive discussions and the responses of each department agent into the consultation results.

[0157] It should be understood that this device corresponds to the above-described embodiment of the method for predicting the risk of cardiac amyloidosis, and is capable of performing the various steps involved in the above-described method embodiment. The specific functions of this device can be found in the description above, and detailed descriptions are appropriately omitted here. The device includes at least one software functional module that can be stored in memory or embedded in the device's operating system (OS) in the form of software or firmware.

[0158] Please see Figure 3The diagram shows a structural schematic of an electronic device provided in an embodiment of this application. An electronic device 300 provided in this application includes a processor 310 and a memory 320. The memory 320 stores machine-readable instructions executable by the processor 310. When the machine-readable instructions are executed by the processor 310, the method described above is performed.

[0159] This application embodiment also provides a computer-readable storage medium 330, on which a computer program is stored. This computer program is executed by a processor 310 to perform the methods described above. The computer-readable storage medium 330 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0160] This application also provides a computer program product, including: a computer program or computer instructions, which are executed by a processor to perform the method described above.

[0161] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0162] It should be understood that the disclosed apparatus and methods can also be implemented in other ways, as provided in the embodiments of this application. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in the flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the accompanying drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending primarily on the functions involved.

[0163] Furthermore, the functional modules of each embodiment in this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. In addition, in the description of this specification, the reference to terms such as "one embodiment," "some embodiments," "example," "specific example," "some examples," etc., means that the specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0164] The above description is only an optional implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of this application should be covered within the protection scope of the embodiments of this application.

Claims

1. A method for predicting the risk of cardiac amyloidosis, characterized by, The method comprises the following steps: obtaining medical information of a suspected patient; processing the medical information by a large language model to obtain auxiliary prediction information containing unsupported point information and triage suggestion results, the unsupported point information being information that is insufficient to support the diagnosis or exclusion of cardiac amyloidosis identified by the large language model; screening department doctors according to the triage suggestion results; sending the auxiliary prediction information to the terminal device of the department doctors according to the triage suggestion results, so as to receive new diagnosis information fed back by the doctors and subsequent medical information of the suspected patient through the terminal device of the department doctors, the new diagnosis information including examination item results supplemented based on the unsupported point information; if the new diagnosis information does not contain exclusion or diagnosis of cardiac amyloidosis risk, processing the subsequent medical information of the suspected patient by the large language model until the unsupported point information in the new diagnosis information no longer prompts the need to supplement the examination item; wherein the new diagnosis information and the auxiliary prediction information are used to exclude or diagnose the cardiac amyloidosis risk; wherein the processing of the medical information by the large language model comprises: extracting key clinical indicators from the medical information and mapping the key clinical indicators to standard term clinical indicators; performing a first retrieval in a diagnosis and treatment guideline knowledge base based on the standard term clinical indicators to obtain a diagnosis and treatment guideline knowledge block; generating a reasoning path and an accurate retrieval question corresponding to the reasoning path by the large language model based on the diagnosis and treatment guideline knowledge block and the medical information; converting the accurate retrieval question into a retrieval embedding vector; performing a second retrieval in a vector database by the retrieval embedding vector to obtain an embedding vector knowledge block with higher relevance to the medical information than the diagnosis and treatment guideline knowledge block, the relevance between the embedding vector knowledge block and the medical information being greater than the relevance between the diagnosis and treatment guideline knowledge block and the medical information; inputting the embedding vector knowledge block and the medical information into the large language model to make the large language model generate the auxiliary prediction information.

2. The method of claim 1, wherein, The screening of department doctors according to the triage suggestion results comprises: determining a range of selectable doctors according to the medical information of the department doctors, the doctors' working hours, the doctors' shift time, holiday markers and outpatient peak indexes; extracting selectable doctor features from each selectable doctor information in the range of selectable doctors; calculating the success rate of each selectable doctor in the medical information based on the selectable doctor features and the medical information using an XGBoost model; screening the department doctors from the range of selectable doctors according to the success rate.

3. The method of claim 2, wherein, After extracting the selectable doctor features from each selectable doctor information in the range of selectable doctors, the method further comprises: adding pseudo-sample features with random rewards to each selectable doctor information in the range of selectable doctors; merging the selectable doctor features and the pseudo-sample features with random rewards to obtain merged doctor features; The merged doctor features are added to the optional doctor features, and the visit success rates of the optional doctors are updated according to the optional doctor features.

4. The method of claim 1, wherein, After the department doctors are screened according to the triage suggestion result, the method further includes: receiving a consultation request sent by a terminal device of the department doctor; starting multi-agent consultation according to the consultation request to obtain a consultation result; sending the consultation result to the terminal device of the department doctor.

5. The method of claim 4, wherein, The multi-agent includes a total agent and department agents; starting multi-agent consultation according to the consultation request includes: parsing a consultation question from the consultation request through the total agent, and determining department agents that need to participate according to the consultation question; answering the consultation question through the department agents to obtain answer results of the department agents; summarizing the answer results of the department agents through the total agent to obtain the consultation result.

6. The method of claim 5, wherein, The answering of the consultation question through the department agents includes: for each department agent of the department agents, querying a professional knowledge block corresponding to the consultation question from a professional knowledge base configured for the department agent; inputting the professional knowledge block corresponding to the consultation question as a context of the department agent into a large language model configured for the department agent; generating an answer result of the department agent through the large language model of the department agent.

7. The method of claim 5, wherein, The summarizing of the answer results of the department agents through the total agent includes: performing consistency checking on the answer results of the department agents through the total agent to obtain a consistency checking result; if the consistency checking result has contradictions, triggering the department agents to interact and discuss through the total agent, and summarizing a final result of the interaction and discussion and the answer results of the department agents into the consultation result.

8. A cardiac amyloidosis risk prediction device, characterized by, The method includes: a medical information acquisition module configured to acquire visit medical information of a suspected patient; an information analysis processing module configured to analyze and process the visit medical information through a large language model to obtain auxiliary prediction information including unsupported point information and a triage suggestion result, the unsupported point information being missing or insufficient information that is insufficient to support a diagnosis or exclusion of cardiac amyloidosis identified by the large language model; a department doctor screening module configured to screen department doctors according to the triage suggestion result; a feedback information receiving module configured to send the auxiliary prediction information to a terminal device of the department doctors according to the triage suggestion result, so as to receive newly diagnosed information and subsequent medical information of the suspected patient fed back by the department doctors through the terminal device of the department doctors, the newly diagnosed information including examination item results supplemented based on the unsupported point information; The risk exclusion diagnosis module is configured to analyze and process subsequent medical information of the suspected patient by the large language model until no check item needs to be supplemented is prompted in the newly added diagnosis information if the newly added diagnosis information does not contain the exclusion or diagnosis of the risk of cardiac amyloidosis; wherein the newly added diagnosis information and the auxiliary prediction information are used to exclude or diagnose the risk of cardiac amyloidosis; The information analysis processing module is specifically configured to extract key clinical indicators from the medical information and map the key clinical indicators to standard term clinical indicators; perform a first retrieval in a diagnosis and treatment guideline knowledge base based on the standard term clinical indicators to obtain a diagnosis and treatment guideline knowledge block; generate a reasoning path and an accurate retrieval question corresponding to the reasoning path by a large language model based on the diagnosis and treatment guideline knowledge block and the medical information; convert the accurate retrieval question into a retrieval embedding vector; perform a second retrieval in a vector database based on the retrieval embedding vector to obtain an embedding vector knowledge block with a higher correlation degree than the diagnosis and treatment guideline knowledge block with the medical information, wherein the correlation degree between the embedding vector knowledge block and the medical information is greater than the correlation degree between the diagnosis and treatment guideline knowledge block and the medical information; and input the embedding vector knowledge block and the medical information into a large language model to enable the large language model to generate the auxiliary prediction information.

9. An electronic device, comprising: Comprising: A processor and a memory, the memory storing machine-readable instructions executable by the processor, the machine-readable instructions being executed by the processor to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program stored on the computer readable storage medium is executed by the processor to perform the method of any one of claims 1 to 7.

11. A computer program product, characterised in that, Comprising: A computer program or computer instructions, the computer program or the computer instructions being executed by the processor to perform the method of any one of claims 1 to 7.

Images