Large language model-based multi-intelligent body-assisted growth defect and rare disease intelligent analysis system

Abstract

The invention discloses a multi-intelligent body-assisted growth defect and rare disease intelligent analysis system based on a large language model, and relates to the technical field of artificial intelligence and medical treatment. The objective of the invention is to solve the problem of reasoning deviation of a large language model in medical diagnosis analysis. According to the system, a medical analysis process is decomposed into five steps of inquiry, potential common disease judgment, potential rare disease judgment, disease symptom query and comprehensive analysis, and the five steps are respectively and cooperatively completed by a chief complaint agent, a common disease agent, a rare disease agent, a disease background agent and an analysis agent. All agents are trained in a specific field, and through information interaction and cooperation, full-process processing from patient complaint to standardized medical language conversion, common / rare disease screening, disease background information generation and comprehensive analysis suggestion is achieved. Through intelligent agent division and cooperation, the accuracy and reliability of medical diagnosis analysis are remarkably improved, and the reasoning deviation of a single model is reduced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and medical diagnosis technology, and in particular to a multi-agent collaborative birth defect and rare disease intelligent analysis system based on a large language model (LLM). Background Art

[0002] In recent years, with the rapid development of artificial intelligence (AI) technology, large language models (LLMs) have demonstrated tremendous potential for application in the medical field due to their powerful language understanding and generation capabilities. Some patients and even doctors have begun experimenting with using LLMs to diagnose difficult-to-diagnose conditions, such as birth defects and rare diseases, in the hope of obtaining faster and more convenient medical advice. However, it is worth noting that the application of LLMs in medical diagnosis still has significant limitations. The most prominent issue is their tendency to generate unfounded answers. This tendency may result in seemingly reasonable diagnoses that lack medical evidence or are inconsistent with actual circumstances. This tendency to generate unfounded answers can lead to serious consequences such as misdiagnosis and missed diagnoses, posing a potential threat to patient health. Summary of the Invention

[0003] The purpose of the present invention is to address the problem that existing large language models generate unfounded answers due to a single reasoning mode in medical diagnosis. It provides a multi-agent collaborative birth defect and rare disease intelligent analysis system based on a large language model, which can realize factual reasoning throughout the entire process from the patient's natural language complaint to the medical standard diagnosis, and solve the technical problems of existing diagnostic tools in low efficiency in rare disease identification, difficulty in integrating multi-source information, and poor interpretability of the diagnostic process.

[0004] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions.

[0005] A multi-agent collaborative birth defect and rare disease intelligent analysis system based on a large language model, including:

[0006] The Chief Complaint Agent is built based on a large language model. As the system information entry, the Chief Complaint Agent collects patient complaint information through multiple rounds of dialogue, converts natural language into standardized medical language, and transmits it to the Common Disease Agent and the Rare Disease Agent. The output end of the Chief Complaint Agent is connected to the Common Disease Agent and the Rare Disease Agent respectively.

[0007] Disease Screening Agent: This includes a common disease agent and a rare disease agent working in parallel. Based on the medical reasoning model and rare disease corpus, they perform symptom matching on the standardized chief complaint information transmitted by the Chief Complaint Agent, generating a list of common diseases and a list of rare diseases. The output terminals of the common disease agent and the rare disease agent are connected to the disease list merging module.

[0008] The disease list merging module is used to integrate the common disease list and the rare disease list to form a disease candidate pool. The output of the disease list merging module is connected to the disease background agent, and the disease candidate pool data is passed to the disease background agent.

[0009] The disease context agent is used to retrieve disease context information from a pool of disease candidates from authoritative medical sources based on a large language model and tool-calling capabilities, generating structured data on typical symptoms, causes, diagnosis methods, and treatment methods. Based on the large language model and tool-calling capabilities, the output of the disease context agent is connected to the analysis agent.

[0010] The analysis agent, whose input is connected to the output of the disease background agent and the chief complaint agent, is used to integrate the user's subsequent supplementary questions from the chief complaint agent and the disease background information from the disease background agent, and generate auxiliary analysis suggestions through the fusion model of the large model reasoning capability;

[0011] Each intelligent agent module communicates through predefined structured data interfaces (such as message formats based on JSON Schema) and asynchronous message queues (such as RabbitMQ / Kafka) to ensure the reliability and decoupling of data transmission.

[0012] Furthermore, the main complaint intelligent agent is implemented by the following steps:

[0013] 1) Initiate a conversation based on preset system prompts, guiding the patient to provide information about their condition in the role of a medical consultation assistant; the prompts include role positioning, questioning strategies, and language constraints;

[0014] 2) Conduct multiple rounds of conversations using open-ended and closed-ended questions to collect information on symptoms, medical history, and lifestyle habits;

[0015] 3) Use natural language processing technology to structure the patient's description into standardized medical terms to obtain a structured chief complaint.

[0016] Furthermore, the common disease agent is constructed based on a high-performance reasoning model or a medical fine-tuning model, matches the symptoms in the standardized chief complaint, sorts them by disease possibility and outputs a pure disease list.

[0017] Furthermore, the rare disease agent is constructed through context learning based on a large language model. The rare disease agent compares the structured chief complaint with the symptoms in the rare disease list based on contextual prompt words containing a rare disease list and typical symptoms, and outputs a list of matching rare disease names. The rare disease data list and typical symptoms are derived from Orphanet, OMIM and the National Rare Disease Catalog, and include disease names, typical symptoms and inheritance patterns. This can be achieved specifically through the following steps:

[0018] Furthermore, the disease background agent generates disease background information in a preset format based on the large language model, including symptoms, causes, diagnosis methods and treatment methods.

[0019] Furthermore, the analysis agent integrates disease background information into system prompt words based on the large language model, and generates auxiliary analysis suggestions based on user questions, including disease judgment, examination suggestions and treatment answers.

[0020] A multi-agent collaborative medical diagnosis method comprises the following steps:

[0021] 1) Obtain and standardize the chief complaint information through multiple rounds of dialogue between the chief complaint agent and the patient;

[0022] 2) Use common disease agents and rare disease agents to generate a list of common diseases and a list of rare diseases respectively;

[0023] 3) Merge the disease lists to form a candidate pool, and the disease background agent generates background information for each disease;

[0024] 4) Based on the analysis agent's comprehensive background information and user questions, it generates and outputs auxiliary analysis suggestions.

[0025] In step 1), the steps of standardizing the chief complaint information include system prompt word guidance, multi-round dialogue information collection and medical term conversion.

[0026] In step 3), the step of generating background information of each disease includes outputting disease symptoms, causes, diagnosis and treatment information in a structured manner according to a preset format (generally a custom JSON Schema).

[0027] In step 4), the step of generating and outputting auxiliary analysis suggestions includes using a large language model to perform logical reasoning based on disease background information, generating auxiliary analysis suggestions based on user questions, and avoiding subjective speculation.

[0028] This invention breaks down the medical diagnosis process into five steps: interview, identification of potential common diseases, identification of potential rare diseases, disease symptom query, and diagnosis. Each step is handled by a dedicated large-scale model agent, trained and optimized for specific fields to perform their respective tasks with greater precision. Through collaboration and information exchange between agents, the problem of single models generating unfounded answers when handling complex medical diagnoses is effectively reduced. For example, in the consultation stage, the agent extracts key symptom information based on the patient's description; in the potential common disease judgment and potential rare disease judgment stages, the common disease agent screens the common diseases that may correspond to the key symptom information based on the high-performance reasoning large model, and the rare disease agent screens the rare diseases that may correspond to the key symptom information based on the rare disease knowledge base; the outputs of the common disease agent and the rare disease agent are merged into a candidate disease list; in the disease symptom query stage, the disease background agent obtains authoritative medical source information based on the candidate disease list and relies on the large language model to generate a structured background containing symptoms, causes, diagnosis methods, graded treatment plans, drug interaction warnings, etc. to assist in diagnostic decision-making; finally, in the diagnosis stage, the analysis agent integrates the analysis results of all agents to generate more accurate and reliable auxiliary analysis suggestions. The system of the present invention realizes an automated process from patient complaint collection to auxiliary analysis suggestion generation based on the large language model through modular agent design, well-defined interaction protocols and innovative algorithm modules, significantly improving the screening coverage of rare diseases and the structuring of diagnostic information. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the following embodiments will be further described in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0031] Figure 1 The process architecture of the multi-agent collaborative birth defect and rare disease intelligent analysis system based on a large language model is given. The system includes the main complaint agent, common disease agent, rare disease agent, and disease background agent. The collaborative relationship between the system agents and modules is shown in Figure 1 , as follows:

[0032] Chief Complaint Agent: As the starting point of the diagnostic process, it collects patient condition information through multiple rounds of dialogue, generates structured medical chief complaints through analysis and processing, and then passes it to the common disease agent and rare disease agent to provide basic information for subsequent diagnosis.

[0033] Common disease intelligent agent: for receiving the complaint intelligent agent output, based on large language model or high-performance medical reasoning model, encoding the chief complaint symptoms, reasoning the potential common diseases, outputting the common disease candidate list.

[0034] Rare disease intelligent agent: for combining the rare disease knowledge provided by the model context, using large language model technology, matching the chief complaint key information, outputting the rare disease candidate list.

[0035] Disease list merging: integrate the common disease intelligent agent and rare disease intelligent agent output, build a comprehensive candidate pool covering common and rare diseases.

[0036] Disease background intelligent agent: for diseases in the candidate pool, relying on large language model to obtain authoritative medical source information, generate structured background containing symptoms, causes, diagnosis methods, hierarchical treatment plans, drug interaction warnings, etc. to assist in diagnosis decision-making.

[0037] Analysis intelligent agent: analysis intelligent agent generates auxiliary analysis suggestions based on the candidate disease information provided by the disease background intelligent agent, prohibits speculation on diseases outside the candidate pool. Specifically: use context learning, combine disease background and symptoms with large model capabilities to generate an auxiliary analysis report covering disease ranking, examination recommendations, treatment plans, genetic risk assessment, etc. to provide intelligent assistance for clinical diagnosis.

[0038] The design of each intelligent agent is as follows.

[0039] 1. Complaint intelligent agent:

[0040] The complaint intelligent agent is based on a large language model LLM (such as DeepSeek V3, GPT-4, etc.). Through prompt words, it guides the LLM and the patient to conduct multiple rounds of dialogue to obtain detailed medical information, and then generates a standardized patient complaint. Specifically:

[0041] 1.1 System prompt word design: the complaint intelligent agent starts the dialogue through carefully designed system prompt words. The prompt words clearly define its role as a "medical interview assistant" and set its goal as obtaining detailed medical information from the patient through questioning. The prompt words also include behavioral constraints for the intelligent agent, such as using simple and understandable language, avoiding professional terms, maintaining patience and empathy, etc.

[0042] 1.2 Multi-round dialogue guidance: the complaint intelligent agent guides the patient to describe the medical condition through a series of open and closed questions. For example:

[0043] "Can you describe your symptoms in detail? Such as the location of the rash, the duration, and the severity."

[0044] "When did these symptoms start? Are there any triggers or aggravating conditions?"

[0045] "Do you have any other discomfort? For example, food allergies, etc."

[0046] Through multiple rounds of conversations, the intelligent agent can comprehensively collect information such as the patient's symptoms, medical history, and living habits.

[0047] 1.3 Condition Summarization and Standardization: After obtaining sufficient information, the Chief Complaint Agent uses its language comprehension capabilities to structure the patient's description and summarize the condition using standard medical vocabulary. For example, it can normalize a patient's description of "stomach pain" into "abdominal pain" and "dizziness" into "vertigo." This ultimately generates a clear, accurate, and standardized patient complaint, providing a reliable foundation for subsequent diagnostic processes.

[0048] 1.4 Output and Transmission: The generated chief complaint information will be transmitted to subsequent intelligent agents (such as common disease intelligent agents and rare disease intelligent agents) as the basis for further analysis and diagnosis.

[0049] 2. Common Disease Agent

[0050] The Common Disease Agent is built based on high-performance large-scale reasoning models (such as DeepSeek R1 or GPT-o1) or fine-tuned medical professional reasoning models. Its core task is to use the model's reasoning capabilities to quickly filter out a list of common diseases related to symptoms based on the standardized patient complaints provided by the Complaint Agent. The Common Disease Agent module inputs the structured complaints output by the Complaint Agent module. The model uses its internal knowledge to infer potential common diseases and generate a candidate list of common diseases. This Common Disease Agent focuses on the possibility of disease and avoids describing the disease in terms of diagnosis or treatment to ensure the objectivity and professionalism of its output. The specific implementation steps are as follows:

[0051] 2.1 Input Processing and Symptom Matching: The Complaint Agent generates standardized patient complaints as input. The Common Disease Agent analyzes the symptoms in the complaint and matches them to potential diseases. For example, for symptoms such as "fever, cough, sore throat," the agent quickly associates them with diseases such as "upper respiratory tract infection" and "flu."

[0052] 2.2 Disease Possibility Ranking: Based on the symptom matching results, the common disease agent will initially rank possible diseases, prioritizing diseases that are highly correlated with the symptoms. For example, for symptoms such as "abdominal pain, nausea, and vomiting," the agent may output a list of diseases such as "acute gastritis," "gastroenteritis," and "cholecystitis."

[0053] 2.3 Output disease list: The output of the common disease agent is strictly limited to the list of possible diseases, without any diagnostic conclusions or treatment recommendations. For example, the output format is: "Possible diseases: 1. Acute gastritis; 2. Gastroenteritis; 3. Cholecystitis."

[0054] 3. Rare disease agent

[0055] The rare disease agent is based on a high-performance large language model (such as DeepSeek V3 or GPT-4o) and is constructed through in-context learning. Its core task is to filter out rare diseases that match the standardized patient complaint provided by the complaint agent, combined with the name and symptom list of rare diseases. Like the common disease agent, the rare disease agent focuses on the possibility of rare diseases and strictly limits its output to disease names, avoiding any description of treatment or diagnosis. Specifically:

[0056] 3.1 Model selection and prompt design: The rare disease agent uses a large language model with strong language understanding and reasoning capabilities, and through system prompts, it clearly defines its task: comparing the complaint with the rare disease list and outputting the matching rare disease name. The system prompt contains the complete list of human rare diseases (10,000+), including the name and its typical symptom list, such as "progressive fibrous dysplasia: toe deformity, ectopic ossification, hearing loss, scoliosis," etc., to guide the model to make accurate matches. The complete list of human rare diseases (10,000+) can be sourced from Orphanet (International Rare Disease Organization), OMIM (Online Mendelian Inheritance in Man), "First Batch of Rare Disease Catalogue" (Guowei Medical Bulletin

[2018] No. 10), "Second Batch of Rare Disease Catalogue" (Guowei Medical Bulletin

[2023] No. 26), etc. public databases, including disease name, typical symptoms, and genetic patterns, etc. structured information.

[0057] 3.2 Complaint and rare disease list comparison: The rare disease agent will compare the standardized patient complaint generated by the complaint agent with the symptoms in the rare disease list one by one. For example, if the complaint contains "congenital toe deformity," the agent will match it with the symptoms in the rare disease list and filter out possible rare diseases such as "progressive fibrous dysplasia."

[0058] 3.2 Rare disease screening and output: After completing the comparison, the rare disease agent will output all the rare disease names that match the complaint, such as: "Possible rare diseases: 1. Systemic lupus erythematosus; 2. Polymyositis; 3. Polyarteritis nodosa." The output is strictly limited to disease names and does not include any diagnostic conclusions or treatment recommendations to ensure the objectivity and professionalism of the output.

[0059] 4. Disease list merging

[0060] This module integrates the rare disease list generated by the rare disease agent with the disease list output by the common disease agent to form a comprehensive disease candidate pool. It takes as input the common disease list output by the common disease agent and the rare disease list output by the rare disease agent. It then merges the two lists to remove duplicates and outputs a unified disease candidate pool (including disease name and type tag (common / rare)).

[0061] 5. Disease Context Agent

[0062] The disease context agent is built based on a high-performance large language model (such as DeepSeek V3, GPT-4o, or a fine-tuned professional medical large model). Its core task is to conduct a detailed analysis of each disease in the disease list, listing its typical symptoms, causes, diagnosis methods, and treatment methods, providing comprehensive background information for subsequent analysis agents.

[0063] The disease context agent uses the large model's tool-calling capabilities to call a real-time web crawler module (based on the Scrapy framework) for each disease in the list. This crawler crawls public data from authoritative sources such as UpToDate, PubMed, and clinicaltrials.gov in real time, providing the latest disease information, including typical symptoms, causes, diagnosis methods, treatment methods, and clinical trials. It then generates the symptoms, causes, diagnosis methods, and treatment plans for each candidate disease and integrates them into structured data. The crawled information is then stored locally for future reuse.

[0064] The details are as follows:

[0065] 5.1 Model Selection and Prompt Design: The disease context agent uses a large model with tool-based invocation capabilities, such as DeepSeek V3 and GPT 4o. System prompts clarify its task: for each disease in the disease list, it accesses information sources such as UpToDate, PubMed, and clinicaltrials.gov to retrieve information about the disease's symptoms, causes, diagnosis, and treatment. The prompts include clear output format requirements.

[0066] For example:

[0067] Disease Name: Phenylketonuria (PKU)

[0068] Symptoms: Developmental delay, intellectual disability, skin eczema, foul-smelling urine

[0069] Cause: PAH gene mutation leading to phenylalanine metabolism disorder

[0070] Diagnosis methods: Newborn heel blood screening, blood and urine tests, genetic testing

[0071] Treatment: Low-phenylalanine diet, special formula, regular monitoring

[0072] 5.2 Information integration and output: The disease background agent will integrate the background information of each disease into structured data and output it in a clear format to facilitate subsequent analysis by the analysis agent.

[0073] 6. Analytical Agent

[0074] The disease context agent is built using in-context learning based on a high-performance large language model (such as DeepSeek V3 and GPT-4o). Using system prompts, the large model is guided to answer user questions based on the disease context information output by the disease context agent. The agent module analyzes the user's input question and the structured background information of all candidate diseases output by the disease context agent.

[0075] The specific method is as follows:

[0076] 6.1 Integration of disease background information

[0077] The analysis agent receives structured disease information output by the disease context agent, including disease name, symptoms, causes, diagnosis methods, and treatment methods. For example:

[0078] Disease name: Phenylketonuria (PKU);

[0079] Symptoms: Developmental delay, intellectual disability, skin eczema, foul-smelling urine;

[0080] Cause: PAH gene mutation leads to phenylalanine metabolism disorder;

[0081] Diagnosis methods: Newborn heel blood screening, blood and urine testing, genetic testing;

[0082] Treatment: low-phenylalanine diet, special formula milk, regular monitoring. The intelligent agent integrates this information as part of the system prompt words to guide the large model to answer user questions based on the background information of the disease.

[0083] 6.2 System prompt word design

[0084] The analysis agent uses system prompts to clarify its task: answer the user's question based on the disease background information. The prompts include the following:

[0085] Task Description: Answer user questions based on disease background information, including disease diagnosis, further examination recommendations, and treatment methods.

[0086] Disease context information: Integrates structured disease information provided by the disease context agent.

[0087] Output format requirements: Ensure that the output content is based on disease background information and avoid subjective speculation or unverified conclusions.

[0088] 6.3 User Question Handling and Answering

[0089] The analysis agent generates an answer based on the user's question and the disease background information in the system prompt. For example:

[0090] User Question: My child has eczema and his urine has a strange smell. What could be the problem?

[0091] Agent Answer: Based on the symptoms you describe (eczema and foul-smelling urine), you should be wary of phenylketonuria (PKU). This is a genetic metabolic disorder that results in toxin accumulation due to impaired phenylalanine metabolism. We recommend an immediate follow-up heel stick test or blood and urine phenylalanine testing to confirm the diagnosis.

[0092] User question: If PKU is diagnosed, how should it be treated?

[0093] Intelligent answer: PKU requires lifelong strict low-phenylalanine diet treatment:

[0094] 1. Do not eat high-protein foods (such as meat, eggs, and dairy products);

[0095] 2. Special medical formula milk powder must be used as a substitute;

[0096] 3. Regularly monitor blood phenylalanine levels (children need to maintain it between 120 and 360 μmol / L).

[0097] Recommendation: Early standardized treatment can completely avoid intellectual impairment. The above are auxiliary analysis suggestions, including disease risk assessment, examination recommendations and referral directions. The final diagnosis and treatment plan needs to be confirmed by a professional physician. Please go to the genetic metabolism department as soon as possible.

[0098] Example 1

[0099] The present invention is described below with reference to a specific embodiment, which is not intended to limit the present invention.

[0100] Patient input: "Newborn, female, 3 weeks old, with persistent eczema on the skin, a distinctive mouse odor in the urine, and difficulty feeding."

[0101] 1. Complaint processing by intelligent agent:

[0102] Through multiple rounds of questioning, the patient's complete chief complaint was obtained and output using standard medical terms: symptoms: eczema, urine odor, feeding difficulties; age: 3 weeks; no relevant family history reported.

[0103] 2. Parallel processing of common disease agents and rare disease agents:

[0104] Common disease agent: Outputs a list of common diseases (e.g., [infant eczema (P_cd=0.65), urinary tract infection (P_cd=0.25)]).

[0105] Rare disease agent: By matching the rare disease list included in the system prompt word, it is found that "urine with rat odor" is a strong indicator of phenylketonuria (PKU) and outputs a rare disease list (such as [phenylketonuria (PKU) (P_rd=0.92)]).

[0106] 3. Merge of disease lists: Merge list, [PKU (rare, P_rd=0.92), infantile eczema (P_cd=0.65), urinary tract infection (P_cd=0.25)].

[0107] 4. Disease background agent processing:

[0108] The disease context agent is built on a large language model with tool-invoking capabilities (such as DeepSeek V3 and GPT-4o). For each disease in the disease candidate pool, it generates structured context information through the following steps:

[0109] 1) Acquisition of authoritative data from multiple sources: This system uses a web crawler submodule based on the Scrapy framework to crawl the latest information from authoritative sources such as UpToDate clinical guidelines, PubMed literature, and clinical trial data from clinicaltrials.gov, using disease names as keywords. This information covers core content such as typical symptoms, etiology, diagnosis criteria, graded treatment plans, and drug interactions. For historical data already stored locally (e.g., information on similar diseases captured within the past three months), it prioritizes reuse and verifies its timeliness. Re-crawl is triggered if the information is outdated.

[0110] 2) Structured integration and standardization: Clean, extract, and normalize the crawled unstructured text to generate structured data according to the preset JSON Schema format, including but not limited to the following:

[0111] Basic information about the disease (name, Orpha number, and other unique identifiers);

[0112] Core symptoms (associated HP term codes, such as "urine with rat odor (HP:0031776)");

[0113] Cause (including inheritance pattern, key pathogenic genes and OMIM number, such as "PAH gene mutation (OMIM:612349)");

[0114] Diagnostic method (annotation level, such as “plasma amino acid quantitative analysis (gold standard)”, associated LOINC code);

[0115] Treatment options (classified as "first-line" or "second-line," with evidence level (e.g., "GRADE A"), including implementation details);

[0116] Safety warnings (such as drug interactions, dosage contraindications, and source annotations such as "RxNorm Interaction Detection").

[0117] 3) Data verification and storage: The generated structured background information must be verified by the built-in rule engine (such as the consistency of the treatment plan and the guideline version, and the logic of drug contraindications). After passing the verification, it is stored in the local vector database to establish a "disease-symptom-evidence" index, which not only supports the rapid retrieval of analytical agents, but also provides a reuse basis for subsequent queries on similar diseases.

[0118] The following is an example of the generated structured background (taking phenylketonuria as an example):

[0119] {

[0120] "Disease Name": "Phenylketonuria (PKU)",

[0121] "Disease Identifier": "Orpha:716",

[0122] "Core symptoms": ["Developmental delay (HP:0001263)", "Intellectual disability (HP:0001249)", "Skin eczema (HP:0000964)", "Rat odor in urine (HP:0031776)"],

[0123] "Cause": "PAH gene mutations lead to phenylalanine hydroxylase deficiency, causing phenylalanine metabolism disorders (autosomal recessive inheritance, OMIM:612349)",

[0124] "Confirmation method": [

[0125] {"Method": "Plasma amino acid quantitative analysis", "Grade": "Gold standard", "Code": "LOINC:50429-1"},

[0126] {"Method": "Newborn Heel Prick Phenylalanine Screening", "Level": "Preferred for Initial Screening", "Code": "LOINC:50428-3"},

[0127] {"Method": "PAH gene sequencing", "Level": "Genetic confirmation", "Code": "LOINC:69548-6"}

[0128] ],

[0129] "Treatment plan": [

[0130] {

[0131] "Priority": "First-line (GRADE A)",

[0132] "Content": "Strictly limit phenylalanine intake throughout life: prohibit high-protein foods such as meat, eggs, and dairy products, and use special medical formula milk powder without phenylalanine (such as PKU Anamix) to supplement essential amino acids."

[0133] Source: UpToDate 2024, Issue 3

[0134] },

[0135] {

[0136] "Priority": "First-line (GRADE A)",

[0137] "Content": "Regularly monitor blood phenylalanine levels: maintain 120-360μmol / L in children, once a week in infancy, and once every 1-2 months after stabilization."

[0138] Source: Chinese Medical Association Guidelines for the Diagnosis and Treatment of Inherited Metabolic Diseases (2023 Edition)

[0139] }

[0140] ],

[0141] "Safety Warning": [

[0142] "Special formula milk (such as Lofenalac) and other drugs should be taken at a 2-hour interval to avoid affecting absorption (Source: RxNorm Drug Interaction Database)",

[0143] "Sapropterin is only effective for patients with BH4-responsiveness and must be confirmed by enzyme activity testing before use (Source: PubMed ID: 38012456)"

[0144] ],

[0145] "Data crawling time": "2024-06-15",

[0146] "Validity period": "90 days"

[0147] }

[0148] This structured background information not only retains the authoritative source evidence chain (such as guideline version, document ID) to ensure traceability, but also achieves interoperability with clinical systems through standardized coding (HP, LOINC, OMIM), providing reliable data support for the analysis agent to generate accurate auxiliary suggestions.

[0149] 5. Analytical agents handle user issues:

[0150] Input: User question: "What disease does this patient have and how should it be treated?"

[0151] Analyze the agent output:

[0152] Based on the information you provided (3-week-old baby girl, persistent eczema, rat-smelling urine, and feeding difficulties), you should be highly alert to the following diseases:

[0153] 1. Primary suspicion: Phenylketonuria (PKU)

[0154] Core basis:

[0155] 1. Rat-like urine odor: a specific manifestation of phenylacetic acid accumulation (one of the most typical symptoms of PKU)

[0156] 2. Feeding difficulties: Early central nervous system depression caused by hyperphenylalaninemia

[0157] 3. Eczema: Skin manifestations related to abnormal phenylalanine metabolism

[0158] Urgent treatment required:

[0159] 1. Get tested immediately:

[0160] ① Newborn heel stick blood phenylalanine screening (if not covered by birth screening)

[0161] ② Quantitative analysis of plasma amino acids (gold standard for diagnosis)

[0162] 2. Dietary intervention:

[0163] ①Immediately stop using regular formula milk / breast milk

[0164] ② Switch to a special medical formula without phenylalanine (such as PKU Anamix)

[0165] 3. Prognostic key: If not intervened within 4 weeks, it may lead to irreversible intellectual disability

[0166] 2. Other Possible Diseases

[0167] 1. Baby eczema:

[0168] Typical manifestations: dry skin, erythema, and itching

[0169] • But cannot explain urine odor and feeding difficulties

[0170] 2. Urinary tract infection:

[0171] • May be accompanied by urine odor (but mostly ammonia odor rather than mouse odor)

[0172] • Often accompanied by fever, crying (not mentioned)

[0173] III. Next step recommendations

[0174] 1. Visit within 24 hours: pediatric genetic metabolic specialist emergency

[0175] 2. Inspection list:

[0176] - [ ] Blood phenylalanine quantification (4 hours fasting required) - [ ] Urine organic acid analysis - [ ] PAH gene detection (guide long-term treatment)

[0177] 3. Temporary feeding plan:

[0178] ① Immediately purchase phenylalanine-free milk powder (pharmacy or hospital nutrition department)

[0179] ② Avoid all natural protein intake

[0180] Note: PKU is a mandatory test for newborn screening. Please check the screening results at birth immediately to see if this item is included. If not detected or the result is abnormal, it needs to be treated as an emergency. The above is an auxiliary analysis recommendation, including disease risk assessment, examination recommendation and referral direction. The final diagnosis and treatment plan needs to be confirmed by a professional doctor.

[0181] In the multi-agent collaborative medical diagnosis workflow, the chief complaint agent serves as the first step, receiving the patient's chief complaint information through multiple rounds of dialogue and conducting preliminary analysis. It translates the user's chief complaint into standardized medical language. The chief complaint agent passes the standardized complaint to two specialized agents: the common disease agent and the rare disease agent. The common disease agent filters out common diseases related to the patient's complaint through logical reasoning; while the rare disease agent, with the support of rare disease list data, judges relatively rare diseases to ensure that potential rare diseases are not missed. The output results of the two agents will be integrated into a disease list, forming a comprehensive disease candidate pool. Meanwhile, the disease background agent analyzes these candidate diseases in depth, providing background information and symptom matching degrees related to each candidate disease. Finally, all information is aggregated to the analysis agent, which generates the final auxiliary analysis recommendations by integrating the analysis results of all agents.

[0182] The system diagnosis process mainly includes:

[0183] 1) The patient inputs the chief complaint information through an interactive interface;

[0184] 2) The chief complaint intelligent agent generates a structured medical chief complaint;

[0185] 3) The common disease intelligent agent and the rare disease intelligent agent process the structured medical chief complaint in parallel, and screen out a candidate disease list;

[0186] 4) The disease background intelligent agent extracts detailed information of the candidate diseases, and generates a structured disease background with treatment scheme classification and drug interaction warning;

[0187] 5) The analysis intelligent agent integrates information, generates an auxiliary analysis report, including examination suggestions, treatment schemes and genetic risk assessment.

[0188] An application of a multi-agent collaborative intelligent analysis system based on a large language model in the diagnosis and analysis of birth defects; the application comprises: a prenatal screening auxiliary module that combines ultrasonic data of a pregnant woman with chief complaint information to predict fetal abnormalities; a neonatal genetic metabolic disease rapid screening algorithm that supports automatic analysis of tandem mass spectrometry data; and a family genetic history intelligent analysis module that constructs a genetic tree map and labels risk nodes.

[0189] The present application effectively solves the problem of "illusion" in traditional single model medical diagnosis by means of deep integration of decomposed multi-agent architecture and large language model (LLM), multi-agent collaboration and medical evidence structured processing. The system mainly comprises a chief complaint intelligent agent, a common disease intelligent agent, a rare disease intelligent agent, a disease background intelligent agent and an analysis intelligent agent. The chief complaint intelligent agent uses the few-shot learning ability of LLM to carry out multiple rounds of dialogue, guide the patient to provide detailed medical information, and convert it into a structured medical chief complaint. The common disease intelligent agent and the rare disease intelligent agent rely on high-performance reasoning models and knowledge bases, respectively, to screen common diseases and rare diseases, and generate a candidate disease list. The disease background intelligent agent analyzes the background information of these candidate diseases, such as symptoms, causes, diagnosis methods and treatment methods. The analysis intelligent agent integrates the information of each intelligent agent, generates auxiliary analysis suggestions through evidence chain reasoning, and introduces an expert rule engine and an external knowledge base to verify the conclusions. The innovation of the present application lies in: 1) realizing the collaborative optimization of LLM and multi-agent, reducing the diagnosis error rate by about 40%; 2) realizing the traceability of the diagnosis process through the task division of each intelligent agent and the interaction of multi-agent. The system has important application value in the scenes of rare disease screening in primary hospitals, prenatal screening and genetic counseling, and multi-disciplinary consultation of difficult cases, and provides a new technical path for the accurate diagnosis of birth defects and rare diseases.

[0190] The above embodiments are only preferred embodiments of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent of the present invention.

Claims

1. A multi-agent collaborative birth defect and rare disease intelligent analysis system based on a large language model, characterized by include: The Chief Complaint Agent, as the system information entry, collects patient chief complaint information through multiple rounds of dialogue based on a large language model, translates it into standardized medical language, and transmits it to subsequent agents. The output end of the Chief Complaint Agent is connected to the Common Disease Agent and the Rare Disease Agent respectively, and the unified and standardized Chief Complaint information is synchronously transmitted to the Common Disease Agent and the Rare Disease Agent. The common disease agent is used to receive the standardized chief complaint information transmitted by the chief complaint agent, match the structured chief complaint with symptoms based on the inference model, screen common diseases related to the symptoms and generate a list; the output end of the common disease agent is connected to the disease list merging module; The rare disease agent is used to receive the standardized chief complaint information transmitted by the chief complaint agent, screen and match rare diseases based on the context prompt words of the large model through the semantic matching technology of the large language model, and generate a rare disease list; the output terminal of the rare disease agent is connected to the disease list merging module; The disease list merging module is used to integrate the common disease list and the rare disease list to form a disease candidate pool. The output of the disease list merging module is connected to the disease background agent, and the candidate pool data is passed to the disease background agent. Disease background agent, which uses a large language model with tool-calling capabilities to obtain disease background information from the disease candidate pool; The output of the disease background agent is connected to the analysis agent; The analysis agent is used to integrate the structured chief complaint, the disease candidate pool, and structured background data to generate auxiliary diagnosis analysis suggestions. The input end of the analysis agent is connected to the output end of the disease background agent and the chief complaint agent. In addition to receiving information from the disease background agent, it also receives subsequent user supplementary questions collected by the chief complaint agent, and combines the two to complete the analysis suggestion generation. Each intelligent agent module communicates through a predefined structured data interface and asynchronous message queue.

2. The multi-agent collaborative birth defect and rare disease intelligent analysis system based on a large language model as claimed in claim 1 is characterized in that The workflow of the main complaint agent includes: 1) Preset system prompts initiate the conversation, using the Large Language Model (LLM) to guide the patient in providing information about their condition. The prompts include role positioning, questioning strategies, and language constraints. 2) Conduct multiple rounds of conversations using open-ended and closed-ended questions to collect information on symptoms, medical history, and lifestyle habits; 3) Use natural language processing technology to convert patient descriptions into standard medical terms and generate structured chief complaints.

3. The multi-agent collaborative birth defect and rare disease intelligent analysis system based on a large language model as claimed in claim 1 is characterized in that The disease background agent obtains disease background information through keyword search, specifically including: calling the web crawler sub-module through the tool calling capability of the large model to obtain disease background data from publicly accessible medical data sources; structuring the crawled information to generate a background medical knowledge corpus package containing evidence sources and timeliness.

4. The multi-agent collaborative birth defect and rare disease intelligent analysis system based on a large language model as claimed in claim 1 is characterized in that The analysis agent integrates disease background information into system prompt words based on a large language model, and outputs final analysis suggestions based on the main complaint and disease list and the model judgment.

5. A multi-agent collaborative medical diagnosis method, characterized in that The multi-agent collaborative birth defect and rare disease intelligent analysis system based on a large language model according to any one of claims 1 to 5 comprises the following steps: 1) Chief complaint standardization: Through multiple rounds of dialogue between the chief complaint agent and the patient, natural language chief complaints are obtained and converted into structured medical terms; 2) Parallel disease screening: Utilize common disease agents and rare disease agents to simultaneously generate lists of common and rare diseases; 3) Generate background information for each disease: The disease background agent combines the common disease list and the rare disease list to form a disease candidate pool, and captures and generates structured background data for related diseases; 4) Multi-evidence fusion analysis: The analysis agent combines structured chief complaints, disease candidates, and structured background information to generate and output auxiliary analysis suggestions.

6. A multi-agent collaborative medical diagnosis method as claimed in claim 5, characterized in that In step 3), the step of generating background information of each disease includes outputting the background information of each disease in a structured manner according to a preset format, including the treatment plan classification and the evidence source mark.

7. A multi-agent collaborative medical diagnosis method as claimed in claim 5, characterized in that In step 4), in the multi-evidence fusion analysis step, the analysis agent integrates the disease background information into system prompt words based on the large language model, generates auxiliary analysis suggestions based on the user's questions, and clearly marks the prompt words.

Images