Intelligent diagnosis method and system for common mental diseases based on multiple agents

Abstract

The invention provides a common mental disease intelligent diagnosis method and system based on multiple agents, and relates to the field of medical artificial intelligence. The method comprises the following steps: S1, extracting diagnosis standards and symptom characteristics of common mental disorders, and constructing a similar patient knowledge base after verification and evaluation of the similar patient knowledge base and expert calibration; s2, acquiring clinical data from a hospital information system, and performing large medical record structuring, clinical scale simplification and scale score analysis on the clinical data to form a similar patient database; and S3, performing symptom matching and scale performance analysis on the input clinical data, constructing a multi-agent mental disease diagnosis framework, and performing multi-agent diagnosis debate based on the multi-agent mental disease diagnosis framework. The technical problems of symptom overlapping and diagnosis subjectivity among common mental diseases are solved by constructing a multi-agent cooperative diagnosis framework, introducing particle size symptom analysis and dynamically integrating structured authoritative medical diagnosis standards.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical artificial intelligence, and more particularly to a common mental illness intelligent diagnosis method and system based on multi-agent. BACKGROUND

[0002] Schizophrenia, bipolar disorder and depression are the three most common mental illnesses, which have become major health and social problems in China. In recent years, the incidence of these three types of mental illness has continued to rise, but due to their complex etiology, diverse symptoms and strong subjectivity, traditional diagnostic methods are highly dependent on the clinical experience of doctors, and there are problems such as low diagnostic consistency and insufficient efficiency. The development of artificial intelligence technology provides a new solution for the auxiliary diagnosis of common mental illnesses.

[0003] In clinical practice, the diagnosis of mental illness is mainly based on clinical medical records (text data); its diagnosis process itself still has many challenges. Although existing intelligent agents based on large language models (LLM) have shown strong decision-making capabilities and have achieved success in general medical diagnosis, there are still significant shortcomings when applied to the diagnosis of common mental illnesses in psychiatry. First, the symptom descriptions (texts) of the three common mental illnesses are highly overlapping, which brings great uncertainty to the diagnosis and differentiation of these three types of diseases. Second, unlike other medical fields, the diagnosis of common mental illnesses in psychiatry lacks objective biomarkers (such as magnetic resonance imaging results, blood tests, and biochemical test results), and mainly relies on subjective symptom-based interviews (such as interviews about hallucinations, delusions, low mood, impulsivity, and irritability), which can lead to an excessive reliance on clinical experience and subjective judgment in the diagnosis process.

[0004] In addition, multi-specialist model-based diagnostic systems can use medical electronic medical record texts to improve diagnostic accuracy through collaborative reasoning between models. However, existing methods still have significant limitations: on the one hand, multi-specialist systems rely on static knowledge bases or pre-trained models, making it difficult to dynamically update medical knowledge, resulting in outdated diagnostic basis; on the other hand, traditional multi-model frameworks lack efficient information retrieval mechanisms, and in the face of complex cases, the decision-making process between specialists may be biased due to insufficient knowledge. Retrieval-Augmented Generation (RAG) has shown advantages in natural language processing, and can enhance the reasoning ability of models by retrieving external knowledge bases in real time. RAG technology provides the possibility for dynamic updating of medical knowledge by models and retrieval of information corresponding to complex cases. Intelligent agent technology combines the advantages of the above multi-specialist and RAG technologies, and is expected to improve the diagnostic accuracy of common mental illnesses. However, this technology has not been fully applied to the field of mental illness diagnosis.

[0005] Therefore, there is an urgent need for a multi-agent mental illness diagnosis framework integrated with a retrieval enhancement mechanism to assist in the accurate diagnosis of common mental illnesses. SUMMARY

[0006] The present application aims to overcome the above technical bottlenecks, dynamically integrate the latest and authoritative medical diagnosis standards, and fully utilize clinical experience to improve the accuracy, explainability, and adaptability of diagnosis, ultimately more accurately and intelligently assisting common mental illness diagnosis.

[0007] According to a first aspect of the present application, a multi-agent based common mental illness intelligent diagnosis method is provided, comprising the following steps: Step S1, extracting the diagnosis standards and symptom characteristics of common mental disorders, and constructing a similar patient knowledge base after verification, evaluation and expert calibration of the similar patient knowledge base; Step S2, obtaining clinical data from a hospital information system and structuring large medical records, simplifying clinical scales and analyzing scale scores to form a similar patient database; Step S3, constructing a multi-agent mental illness diagnosis framework, embedding the similar patient knowledge base and the similar patient database into the multi-agent mental illness diagnosis framework, increasing user input, and conducting multi-agent diagnosis debate.

[0008] On the basis of the above technical solution, the present application can also be improved as follows.

[0009] Optionally, the common mental disorders include but are not limited to common mental disorders such as schizophrenia, bipolar disorder and depression; the extraction of diagnosis standards and symptom characteristics of common mental disorders includes: The diagnosis standards and symptom characteristics are extracted from the authoritative guide DSM-5 in the field of mental illness diagnosis, the complex differential diagnosis content is deconstructed into clear discrimination indicators by GPT-4o, and these discrimination indicators are converted into discrete structured units as the external knowledge base of the intelligent agent.

[0010] Optionally, the similar patient knowledge base is constructed after verification, evaluation and expert calibration of the similar patient knowledge base includes: A multi-level verification, evaluation and expert calibration mechanism is introduced to comprehensively evaluate the accuracy and integrity of all extracted diagnosis standards and symptom descriptions in the knowledge base, and to conduct a key review of the entries of core mental illnesses covering emotional disorders; in this process, any found inaccuracies, ambiguities or omissions are manually reviewed and revised by professional medical experts.

[0011] Optionally, the large medical record structured processing of the clinical data comprises: extracting key clinical elements from the anonymized hospital database to comprehensively capture patient background information; removing explicit disease labels in the medical history and converting absolute dates into relative time expressions, while ensuring that the patient's true symptoms are not tampered with; integrating the preliminarily extracted structured elements with the processed medical history data to form a coherent and consistent contextual understanding; reorganizing the integrated medical records into a standardized structured format.

[0012] Optionally, the clinical scale simplification of the clinical data comprises: Based on the analysis of common psychological scales, by calculating the score of each scale question and the Pearson correlation between the total score and the presence of emotional disorders, the top 5% of key scale items with the highest correlation are quantitatively screened out; Integrating expert clinical considerations, if the correlation score of a scale question with important clinical significance is slightly lower than the statistical threshold, it will still be strategically included in the diagnostic input; Finally, the quantitative score is converted into an interpretable natural language description.

[0013] Optionally, the scale score analysis of the clinical data comprises: The numerical score of the patient on the preferred scale question is converted into a coherent text description paragraph according to the pre-defined rules and question content and option description.

[0014] Optionally, before constructing the multi-agent mental illness diagnosis framework, it further comprises: symptom matching and scale performance analysis on the input clinical data; wherein, Symptom matching: systematically comparing the user-input clinical symptoms with the DSM-5 diagnostic criteria in the constructed similar patient knowledge base to identify potential mood disorder characteristics; Scale performance analysis: evaluating the user-input scale data, quantifying symptom severity, and providing objective evidence for the agent's diagnostic reasoning.

[0015] Optionally, the multi-agent mental illness diagnosis framework comprises: Angel.R mode agent: does not refer to historical case data in the similar patient database during diagnosis, mainly relies on granular symptom analysis results, user-input structured medical records and simplified scale data for reasoning; Angel.D mode agent: references similar historical cases retrieved from the similar patient database as contextual information, but does not perform in-depth analysis to assist in diagnosis; Angel.C mode intelligent agent: detailed comparison and analysis of similar historical cases retrieved from the similar patient database, and extraction of insights from historical experience.

[0016] Optionally, after the multi-agent diagnosis debate, the method further comprises: Outputting the final diagnosis result and the detailed reasons supporting the judgment, and the output result includes: clear diagnosis conclusion, detailed reasoning process and supporting evidence.

[0017] According to the second aspect of the present application, a multi-agent-based common mental illness intelligent diagnosis system is provided, comprising: A similar patient knowledge base construction module is used to extract the diagnostic criteria and symptom characteristics of common mental disorders, and the similar patient knowledge base is constructed after verification, evaluation and expert calibration of the similar patient knowledge base; A clinical data collation and analysis module is used to obtain clinical data from a hospital information system, and to perform large medical record structuring, clinical scale simplification and scale score analysis on the clinical data, forming a similar patient database; A multi-agent diagnosis module is used to construct a multi-agent mental illness diagnosis framework, embed the similar patient knowledge base and the similar patient database into the multi-agent mental illness diagnosis framework, increase user input, and perform multi-agent diagnosis debate.

[0018] The technical effects and advantages of the present application are: The present application provides a multi-agent-based mental illness intelligent diagnosis method and system. The method and system aim to solve the core challenges of symptom overlap and subjective diagnosis among common mental illnesses. It optimizes the existing psychological scale evaluation system by constructing a multi-agent collaborative diagnosis framework, introduces granular symptom analysis, dynamically integrates structured authoritative medical diagnosis standards (DSM-5) and a searchable similar patient knowledge base constructed from historical clinical data. It overcomes the information asymmetry and subjectivity problems in mental illness diagnosis, and ultimately provides more accurate and intelligent assistance for the diagnosis of common mental illnesses, providing more persuasive, data-supported and interpretable diagnosis support. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A multi-agent-based common mental illness intelligent diagnosis method flowchart is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of the present application.

[0021] In the prior art, the three common mental diseases of schizophrenia, bipolar disorder and depression have high morbidity and great social harm, and have become a prominent problem affecting public health. The diagnosis of common mental diseases highly depends on the subjective experience of clinical experts, lacks objective biological markers, and the three diseases have many overlapping symptoms, which makes it difficult to accurately diagnose them.

[0022] It can be understood that based on the defects in the background art, the embodiments of the present application propose an intelligent diagnosis method for common mental diseases based on multi-agent, specifically as shown in Figure 1 The method comprises the following steps: Step S1, extract the diagnostic criteria and symptom characteristics of common mental disorders, and construct a similar patient knowledge base after verification, evaluation and expert calibration of the similar patient knowledge base; It should be noted that the common mental disorders described in the embodiments of the present application include but are not limited to 18 common mental disorders such as schizophrenia, bipolar disorder, depression, etc.

[0023] The extraction of the diagnostic criteria and symptom characteristics of common mental disorders includes: extracting the diagnostic criteria and symptom characteristics from the authoritative guide in the field of mental disease diagnosis, i.e. "Diagnostic and Statistical Manual of Mental Disorders (Fifth Edition)" (DSM-5, American Psychiatric Association, etc., 2013).

[0024] In order to obtain professional diagnostic criteria, the above-mentioned "Diagnostic and Statistical Manual of Mental Disorders (Fifth Edition)" (DSM-5, American Psychiatric Association, etc., 2013) systematically describes the classification criteria and diagnostic specifications of mental disorders, and describes the symptom characteristics, diagnostic points and related clinical manifestations in detail. As an important tool in the field of psychiatry, DSM-5 ensures the consistency and accuracy of mental health diagnosis, and is an indispensable reference for clinical practice and scientific research. The present application converts these authoritative standards into discrete structured units as an external knowledge base of the agent.

[0025] Specifically, the present application extracts the diagnostic criteria and symptom characteristics of 18 common mental disorders such as schizophrenia, bipolar disorder, depression, etc., which fully covers all core symptom groups of "schizophrenia spectrum and other psychotic disorders", "bipolar and related disorders", "depressive disorders" and other core symptom groups. In order to improve the simplicity and clarity of the expression, and to facilitate the comparison with the patient's medical record, the symptom items in DSM-5 (usually listed independently) can be restructured and optimized using a large language model (such as GPT-4o), which can be converted into a format that is easier for machines to process and understand, while ensuring the independence and integrity of the diagnostic criteria.

[0026] At the same time, the complex differential diagnosis content is deconstructed into clear discriminant indicators using a large language model, which can accurately distinguish the clinical manifestations of the three major common mental diseases from other mental diseases.

[0027] To ensure that the constructed similar patient knowledge base has the highest clinical accuracy and practicality, the verification, evaluation and expert calibration of the similar patient knowledge base include: A multi-level verification, evaluation and expert calibration mechanism is introduced to comprehensively evaluate the accuracy and integrity of all extracted diagnostic criteria and symptom descriptions in the knowledge base, especially the core mental illness items covering emotional disorders (such as depression, bipolar disorder). In this process, any inaccuracies, ambiguities or omissions found are manually reviewed and revised by professional medical experts.

[0028] The specific process of manual review and revision is as follows: Symptom description accuracy verification: all extracted diagnostic criteria in the knowledge base are audited for clinical compliance to ensure consistency with the latest medical consensus and practice; diagnostic content integrity test: strictly check whether the key diagnostic elements in the knowledge base are missing to ensure comprehensive coverage of knowledge; expert manual calibration: all expression deviations, content omissions or inconsistencies with clinical practice found in the above verification process are manually reviewed and calibrated by a team of professional clinical experts. This key step ensures that the knowledge base not only comes from authoritative literature, but also incorporates the meticulous insights and experience of clinical experts, thereby achieving the highest professional standards and providing a solid and reliable knowledge base for subsequent intelligent diagnosis. By using the above method, the embodiments of the present application not only ensure the high accuracy of the diagnostic knowledge, but also effectively handle the symptom overlap problem commonly seen in different mental diseases, and overcome the limitations of traditional knowledge bases in handling complex and subjective cases, so that the diagnostic assistance system can provide more persuasive and explainable judgments.

[0029] Step S2, obtaining clinical data from a hospital information system, and performing large medical record structuring, clinical scale simplification, and scale score analysis on the clinical data to form a similar patient database; To simulate the diagnostic expertise of psychiatrists and enable the agent to effectively reason about clinical experience knowledge, the present application integrates heterogeneous and dispersed raw patient clinical data obtained from a hospital information system into the agent's retrieval enhancement framework. These carefully processed data build a rich retrieval database, enabling the agent to efficiently identify similar cases and their corresponding diagnostic results, thereby effectively simulating and utilizing clinical experience knowledge. By combining structured diagnostic criteria with real case data, the framework of the present application bridges the gap between computer decision-making and the subtle clinical understanding required for precise psychiatric assessment.

[0030] The large medical record structuring of the clinical data includes: In view of the challenges of information dispersion, redundancy, or sensitivity commonly existing in raw medical records, the present application designs a systematic medical record processing procedure. This procedure aims to ensure data security, enhance temporal reasoning ability, and optimize symptom extraction to facilitate accurate analysis by the LLM-based agent.

[0031] Specifically, the medical record processing procedure includes: Raw data extraction: Extract key clinical elements such as gender, age, occupation, visit date, chief complaint, and present illness history from anonymized hospital databases to comprehensively capture patient background information; Present illness history processing: To prevent data leakage and enhance temporal reasoning, remove explicit disease labels from the present illness history and convert absolute dates to relative time expressions (e.g., "3 months ago"), while ensuring that patient true symptoms are not tampered with; Key element integration: Integrate the preliminarily extracted structured elements with the processed present illness history data to form a coherent and consistent contextual understanding; Medical record structuring: Reorganize the integrated medical records into a standardized structured format (JSON) containing key symptoms and background information. This processing enables efficient identification and matching of diagnosis-related symptoms scattered throughout the original records, significantly improving the efficiency and accuracy of subsequent analysis.

[0032] To ensure that the scale data received by the agent is highly relevant and free of redundant noise, the present application optimizes and simplifies the psychological scale evaluation system for patients.

[0033] The clinical scale simplification of the clinical data includes: Data-driven question screening: Based on the analysis of common psychological scales, the top 5% of key scale items with the highest correlation between the score of each scale question (and total score) and the presence of emotional disorders are quantified by calculating the Pearson correlation. These statistically verified questions naturally cluster into core symptom groups such as depressive mood, loss of interest, anxiety, insomnia, and suicidal tendencies, ensuring the high effectiveness and specificity of the diagnostic input; Clinical calibration and comprehensive enhancement: To ensure the comprehensiveness and clinical delicacy of the diagnosis process, the invention integrates expert clinical considerations. Even if the correlation score of some important PHQ-9 scale questions (such as PHQ9_Q2 "depressive mood" and PHQ9_Q1 "loss of interest") is slightly lower than the statistical threshold, they will still be strategically included in the diagnostic input. This preferred strategy, which combines statistical correlation analysis and clinical expert insight, overcomes the limitations of traditional scale assessment, which relies too much on total scores and ignores individual item details, significantly reducing diagnostic noise and ensuring that the system can focus on core symptom indicators that are most discriminative and clinically valuable for emotional disorder diagnosis; Natural language analysis of scale scores: To enhance the intelligent agent's understanding and reasoning ability for patient scale performance, the invention converts quantitative scores into interpretable natural language descriptions.

[0034] Scale score analysis of clinical data includes: Semantic conversion: Convert the numerical score of the patient on the preferred scale question described above into a coherent text description paragraph according to pre-defined rules and question content and option descriptions. For example, a PHQ-9 scale score of "2" will be converted to "more than half the time in the past two weeks feeling low, depressed or hopeless."

[0035] Through the above semantic conversion method, the intelligent agent can analyze scale data in a way that is closer to the understanding of human clinicians, overcoming the semantic ambiguity that may exist in pure numerical scores, thereby improving the effectiveness of the analysis and the interpretability of the diagnosis.

[0036] Finally, through the above refinement, the embodiment of the invention can efficiently extract, integrate and understand key information from multiple sources of heterogeneous clinical data, providing a solid and high-credibility data foundation for subsequent intelligent diagnosis of the multi-agent framework.

[0037] Step S3, constructing a multi-agent mental illness diagnosis framework, embedding similar patient knowledge base and similar patient database into the multi-agent mental illness diagnosis framework, adding user input, and conducting multi-agent diagnosis debate.

[0038] When receiving real-time clinical information from the user, first analyze the user's real-time clinical information for symptom matching and scale performance analysis to generate the necessary basis for intelligent agent diagnosis.

[0039] Specifically, before constructing the multi-agent mental illness diagnosis framework, the input clinical data is subjected to symptom matching and scale performance analysis, wherein, Symptom Matching: The user-input clinical symptoms are systematically compared with the DSM-5 diagnostic criteria in the step-constructed similar patient knowledge base to identify potential mood disorder characteristics. This process uses a high-dimensional semantic embedding model (BGE-M3) to calculate the semantic similarity between symptoms and diagnostic criteria, providing the agent with a quantitative symptom matching result.

[0040] Scale Performance Analysis: The user-input scale data (including the simplified scale and natural language analysis results) are evaluated to quantify the symptom severity, providing an objective basis for the agent's diagnostic reasoning.

[0041] Then, the multi-agent mental illness diagnosis framework is constructed.

[0042] It should be noted that the multi-agent mental illness diagnosis framework (Retrieval-augmented Multi-agent Framework) provided by the embodiment of the present application receives real-time clinical information (including medical records and scale data) from the user, combines the verified and calibrated similar patient knowledge base constructed in step S1 and the similar patient database formed after processing in step S2, and constructs and structured debates through the collaboration of agents to achieve high-precision and interpretable diagnostic assistance. This framework aims to overcome the "hallucination" problem that may occur in a single agent and the limitations of knowledge retrieval, and to construct an efficient and reliable diagnosis process.

[0043] To achieve multi-perspective and dynamic diagnostic reasoning, the present application constructs multiple diagnostic agents according to three modes (Angel.R, Angel.D, Angel.C) and collects them to form the core "CogniCore Diagnostic Swarm". These agents use the above-mentioned pre-processed user real-time clinical information, the constructed similar patient knowledge base, and the similar case information obtained from the similar case retrieval and distribution mechanism in the constructed similar patient database to independently generate preliminary diagnostic analysis and reasoning.

[0044] The multi-agent mental illness diagnosis framework includes: Angel.R mode agent: does not refer to historical case data in the similar patient database during diagnosis, mainly relying on granular symptom analysis results, user-input structured medical records and simplified scale data for reasoning; Angel.D mode agent: reference similar historical cases retrieved from the similar patient database as contextual information, but do not conduct in-depth analysis, assist in diagnosis; Angel.C mode agent: conduct detailed comparison and analysis on similar historical cases retrieved from the similar patient database, thereby refining insights from historical experience.

[0045] To effectively utilize the clinical experience of similar cases, a BGE-M3 high-dimensional semantic embedding model is used, and its multi-language semantic coding capability ensures high accuracy of the retrieval results in the diversity of symptom descriptions and cross-language case matching. The system selects the MxN cases with the highest similarity from the similar patient database constructed in step S2 (M is the number of agents in the intelligent nuclear cooperative diagnosis group, and N is the number of cases allocated to each agent), and distributes them to the M agents in the intelligent nuclear cooperative diagnosis group in a distributed manner.

[0046] Further, to solve the possible diagnosis disagreement in the intelligent nuclear cooperative diagnosis group and further improve the robustness and explainability of the final diagnosis decision, the invention embodiment designs a positive argument agent (Positive Agent: based on the preliminary diagnosis result of the intelligent nuclear cooperative diagnosis group, systematically traverses and supports all possible mental illness diagnosis conclusions, and provides detailed arguments and supporting evidence for each conclusion), a negative argument agent (Negative Agent: questions and challenges the diagnosis conclusion proposed by the positive side, and proposes refutation opinions from the aspects of symptom overlap, potential differential diagnosis possibility, inconsistency or insufficient evidence in the data, etc.), and a judge agent (Judge Agent: as a neutral judge, summarizes the arguments of the positive and negative sides and the diagnosis conclusions of the intelligent nuclear cooperative diagnosis group. Its core responsibility is to evaluate the quality of reasoning, the sufficiency of evidence, and the rigor of logic of each party. Through an iterative reasoning and data sharing mechanism, the system systematically explores multiple diagnosis possibilities, significantly improves the accuracy and explainability of diagnosis, and realizes the robust handling of complex cases.

[0047] Each agent in the intelligent nuclear cooperative diagnosis group generates an independent diagnosis conclusion and reasoning path. The judge agent integrates the output results of each agent. If consensus is reached, the judge agent directly outputs the final diagnosis report and the basis; if there is a disagreement, multiple rounds of debate are started until a consistent conclusion is reached or the preset upper limit of the number of debate rounds is reached. This structured dynamic debate architecture deeply collaborates with multiple agent variants, systematically explores multiple diagnosis possibilities through an iterative reasoning and real-time data sharing mechanism, significantly improves the accuracy and explainability of diagnosis, and realizes the robust handling of complex cases.

[0048] Finally, after the multi-agent diagnosis debate, it also includes: debate result report output, output the final diagnosis result and the detailed reasons supporting the judgment.

[0049] Specifically, after multiple rounds of efficient and structured debate, when the decision-making agent believes that sufficient consensus has been reached or that further debate cannot improve decision quality, the final diagnostic result and detailed reasons supporting the judgment will be output. This report not only contains clear diagnostic conclusions (yes / no emotional disorder, what kind of emotional disorder), but also provides detailed reasoning processes (clearly showing the decision-making path of the agent from input data to the final diagnosis); supporting evidence (listing all key data points, scale performance, knowledge base references, and similar case analysis supporting the diagnosis).

[0050] The output of this report significantly enhances the credibility and explainability of the diagnostic results, providing clinicians with more decision-making value and transparent auxiliary information, bridging the gap between AI diagnosis and clinical practice.

[0051] In order to more clearly illustrate the technical solutions provided by the present application, a hypothetical clinical diagnosis scenario is used as an example for detailed description below.

[0052] Specific implementation examples Scenario description: A patient, Zhang San (using a pseudonym for privacy protection), presents with "emotional depression, insomnia for 3 months". The system of the present application has completed steps S1 and S2 (construction of knowledge base and database) in advance. When receiving real-time clinical information (including medical history and scale data) of the patient, it immediately starts the intelligent diagnosis process of step S3.

[0053] Diagnosis basis generation: After the system receives the real-time clinical information of patient Zhang San, it first performs scale performance analysis and symptom matching to generate the basis required for intelligent agent diagnosis.

[0054] Scale performance analysis: The system evaluates the simplified scale data and natural language analysis results input by Zhang San.

[0055] Symptom matching: The system extracts clinical symptoms from Zhang San's medical history, such as "emotional depression", "loss of interest", "difficulty falling asleep", etc. Using a high-dimensional semantic embedding model (BGE-M3), these symptoms are systematically compared with the pre-set DSM-5 diagnostic criteria in the knowledge base.

[0056] Similar case retrieval and distribution: The system uses the BGE-M3 model to retrieve the M×N historical cases with the highest similarity to Zhang San's case from the "similar patient database". These cases are distributed to the diagnosis intelligent agents (Angel.R, Angel.D, Angel.C mode combination, M is the total number of Angel.D and Angel.C mode intelligent agents) in the intelligent nuclear cooperative diagnosis group as reference information.

[0057] "Intelligent nuclear cooperative diagnosis group" independently generates preliminary diagnosis conclusion and reasoning path: Each intelligent agent in the intelligent nuclear cooperative diagnosis group (for example, the total number of intelligent agents is 3, M = 2, including an Angel.R, an Angel.D, and an Angel.C) uses its preprocessed user real-time clinical information, knowledge base, and assigned information in the similar patient database to independently generate its preliminary diagnosis analysis and reasoning. Angel.R mode intelligent agent A: only according to symptom matching and scale analysis, it is preliminarily judged as major depressive disorder, and it is believed that the core symptom group has a high aggregation degree (the detailed discussion is omitted...); Angel.D mode intelligent agent B: combined with similar case reference, it is preliminarily judged as major depressive disorder, and it is believed that the high similarity with historical cases supports this diagnosis (the detailed discussion is omitted...); Angel.C mode intelligent agent C: through detailed comparative analysis, it is preliminarily judged as major depressive disorder, but it proposes to identify the possibility of generalized anxiety disorder, the reason being that there is symptom overlap and some similar cases show complexity.

[0058] Structured diagnosis debate: To solve the possible diagnosis disagreement within the intelligent nuclear cooperative diagnosis group (for example, the identification requirement proposed by Angel.C above), and further improve the robustness and explainability of the final diagnosis decision, the system immediately starts the diagnosis debate process composed of a pro argument intelligent agent, a counter argument intelligent agent, and a ruling intelligent agent.

[0059] Argument construction and challenges: Pro argument intelligent agent (Positive Agent): Based on all preliminary different diagnosis results in the intelligent nuclear cooperative diagnosis group (as well as the inclination of the intelligent nuclear cooperative diagnosis group), all possible mental illness diagnosis conclusions are systematically traversed and supported.

[0060] Core argument: In this case, the clinical manifestations of patient Zhang San are highly indicative of major depressive disorder.

[0061] Detailed argument: The patient's report of "loss of interest in previously loved basketball and games" is a typical anhedonia, which is one of the core symptoms of major depressive disorder in the DSM-5 diagnostic criteria, with high diagnostic specificity and significant differentiation from general anxiety. It is explicitly stated that the severity and persistence of the symptom. The patient also has persistent "mood swings", "lack of energy", and "occasional thoughts of 'life is not worth living'". These symptoms are interrelated and collectively constitute the core symptom group of depressive disorder, and meet the course criteria of DSM-5. The symptoms have led to significant reduction in work efficiency and loss of interest, meeting the functional impairment criteria for major depressive disorder. In the similar patient database, multiple historical cases with clinical manifestations similar to Zhang San's (especially anhedonia and persistent mood swings) were eventually diagnosed as major depressive disorder, providing strong reference evidence.

[0062] Negative Agent: The Negative Agent challenges and questions the positive side's diagnostic conclusion, and puts forward refutation opinions from the aspects of symptom overlap, potential differential diagnosis, inconsistency in data, or insufficient evidence, etc.

[0063] Core argument: Although patient Zhang San has symptoms of depression, it is still necessary to further identify whether it is anxiety as the core performance or whether the depressive symptoms are secondary.

[0064] Detailed refutation: Symptom overlap analysis: The patient complains of "insomnia" and "irritability", which are also very common in generalized anxiety disorder. Although there is emotional depression, the first symptom may be anxiety related to sleep problems rather than core depressive mood. The differential criteria for such symptoms in DSM-5 need to be carefully considered.

[0065] Judgment and iterative reasoning: The Judge Agent as a neutral judge summarizes and evaluates all the information from the three parties (the independent diagnostic conclusions and reasoning paths of the intelligent agents in the intelligent nuclear cooperative diagnosis group, the positive side's arguments, and the negative side's challenges). If the output results of each intelligent agent have reached a consensus, the Judge Agent will directly output the final diagnosis report. If there is a difference, the system will start multiple rounds of debate (such as this example), until it converges to a consistent conclusion or reaches the upper limit of the number of debate rounds.

[0066] Debate result report output: The system outputs the final diagnosis report, which includes the clear diagnosis conclusion (major depressive disorder), detailed reasoning process, and supporting evidence (lists all key data points supporting the diagnosis, such as "similar case retrieval results support the diagnosis of depression", and clearly states that "the debate process has fully considered and ruled out the possibility of generalized anxiety disorder as the main diagnosis").

[0067] In summary, the embodiments of the present application have the following technical effects: Firstly, the embodiments of the present application propose a retrieval-enhanced multi-agent framework for emotional disorder diagnosis: the framework is composed of multiple specialized diagnostic agents, each of which independently performs a diagnostic process and uses different degrees of historical case dependence to balance historical data and individual differences. The diagnostic opinions of the intelligent agents are integrated through a structured debate mechanism to form the final judgment. This helps to solve the difference of opinions, improve the robustness and explainability of the diagnosis.

[0068] Secondly, the present application rethinks and optimizes the existing psychological scale evaluation system through a mixed intelligent feature selection method. The core is to systematically calculate the Pearson correlation between the scores of each major psychological scale question (including the total score) and the presence or absence of emotional disorders, thereby accurately selecting the top 5% of key scale items with the highest correlation. These statistically verified questions naturally cluster into core symptom groups such as depressive mood, loss of interest, anxiety, insomnia, and suicidal tendencies, significantly improving the efficiency and specificity of diagnostic input. Secondly, to ensure the comprehensiveness and clinical delicacy of the diagnosis process, the present application strategically integrates scale questions with important clinical significance (such as specific PHQ-9 items), even if their correlation scores are slightly lower than the statistical threshold, they will be included based on expert clinical considerations. This combination of statistical correlation analysis and clinical expert insight optimizes the strategy, overcoming the limitations of traditional scale evaluation that relies too much on total scores and ignores individual item details, effectively reducing diagnostic noise and ensuring that the system can focus on core symptom indicators that are most discriminative and clinically valuable for emotional disorder diagnosis, laying a high-quality foundation for subsequent granular analysis.

[0069] Finally, beyond traditional scale total score evaluation, the present application conducts detailed analysis at the item level to capture and understand subtle symptom patterns and individual differences within the scale, achieving more personalized and accurate evaluation. Dynamically integrating structured authoritative medical diagnostic criteria (DSM-5) and a searchable similar patient knowledge base constructed from historical clinical data provides the intelligent agent with rich and accurate contextual references, making up for the lack of data and privacy protection in real clinical practice.

[0070] Additionally, according to the second aspect of the present application, an intelligent diagnosis system for common mental disorders based on multi-agent is provided, comprising: a similar patient knowledge base construction module for extracting the diagnostic criteria and symptom characteristics of common mental disorders, and constructing a similar patient knowledge base after verification, evaluation and expert calibration of the similar patient knowledge base; a clinical data sorting and analysis module for obtaining clinical data from a hospital information system and performing large medical record structuring, clinical scale simplification and scale score analysis to form a similar patient database; a multi-agent diagnosis module for constructing a multi-agent mental disorder diagnosis framework, embedding the similar patient knowledge base and similar patient database into the multi-agent mental disorder diagnosis framework, increasing user input, and conducting multi-agent diagnosis debate.

[0071] It can be understood that the common mental illness intelligent diagnosis system based on multi-agent provided by the present application corresponds to the common mental illness intelligent diagnosis method based on multi-agent provided by the foregoing embodiments, and the related technical features of the common mental illness intelligent diagnosis system based on multi-agent can refer to the related technical features of the common mental illness intelligent diagnosis method based on multi-agent, which will not be described here.

[0072] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for limiting the present application, although the foregoing embodiments of the present application are described in detail, for those skilled in the art, it still can be modified, or part of the technical features of the equivalent replacement, within the spirit and principles of the present application, any modification, equivalent replacement, improvement, etc., should be included within the scope of the present application.

Claims

1. A multi-agent based intelligent diagnosis method for common mental illnesses, characterized by: The following steps are involved: Step S1: extract the diagnostic criteria and symptom characteristics of common mental disorders, and construct a similar patient knowledge base after verification, evaluation and expert calibration; Step S2: Acquire clinical data from the hospital information system, structure the clinical data into large medical records, simplify the clinical scales, and analyze the scale scores to form a similar patient database; Step S3: construct a multi-agent mental illness diagnosis framework, embed the similar patient knowledge base and similar patient database into the multi-agent mental illness diagnosis framework, increase user input, and conduct multi-agent diagnostic debate.

2. The multi-agent-based intelligent diagnosis method for common mental illnesses according to claim 1, characterized in that: The common mental disorders include but are not limited to schizophrenia, bipolar disorder and depression; The extracted diagnostic criteria and symptom characteristics of common mental disorders include: The diagnostic criteria and symptom characteristics are extracted from DSM-5, the authoritative guide for the diagnosis of mental illnesses. GPT-4o is used to deconstruct the complex differential diagnosis content into clear discriminant indicators, and these discriminant indicators are converted into discrete structured units as the external knowledge base of the intelligent agent.

3. The multi-agent-based intelligent diagnosis method for common mental illnesses according to claim 1, characterized in that: The construction of a similar patient knowledge base after verification, evaluation and expert calibration of the similar patient knowledge base includes: A multi-level verification, evaluation and expert calibration mechanism was introduced to conduct a comprehensive accuracy and completeness assessment of all extracted diagnostic criteria and symptom descriptions in the knowledge base, and to focus on reviewing entries covering core mental illnesses covering mood disorders. During this process, any inaccuracies, ambiguities or omissions found were manually reviewed and revised by professional medical experts.

4. The multi-agent-based intelligent diagnosis method for common mental illnesses according to claim 1, characterized in that: Structuring clinical data into large medical records includes: Extract key clinical elements from anonymized hospital databases to comprehensively capture patient background information; Remove explicit disease labels from the current medical history and convert absolute dates into relative time expressions, while ensuring that the patient's true symptoms are not tampered with; Integrate the initially extracted structural elements with the processed history of present illness data to form a coherent and consistent contextual understanding; Reorganize consolidated medical records into a standardized, structured format.

5. The multi-agent-based intelligent diagnosis method for common mental illnesses according to claim 1, characterized in that: Clinical scale simplification of clinical data includes: Based on the analysis of common psychological scales, the Pearson correlation between the score of each scale question and the total score and the presence of mood disorders was calculated to quantitatively screen the top 5% of key scale items with the highest correlation; Integrate expert clinical considerations and strategically include clinically important scale questions in the diagnostic input if their relevance scores are slightly below the statistical threshold; Finally, the quantitative scores are converted into interpretable natural language descriptions.

6. The multi-agent-based intelligent diagnosis method for common mental illnesses according to claim 1, characterized in that: Scale score analysis of clinical data includes: The patient's numerical scores on the preferred scale questions are converted into coherent text description paragraphs based on predefined rules and question content and option descriptions.

7. The multi-agent-based intelligent diagnosis method for common mental illnesses according to claim 1, characterized in that: Before constructing the multi-agent mental illness diagnosis framework, the following steps are also included: The input clinical data were analyzed for symptom matching and scale performance; Symptom matching: Systematically compares the clinical symptoms entered by the user with the DSM-5 diagnostic criteria in the constructed knowledge base of similar patients to identify potential mood disorder characteristics; Scale performance analysis: Evaluate the scale data entered by the user, quantify the severity of symptoms, and provide an objective basis for the diagnostic reasoning of the intelligent agent.

8. The multi-agent-based intelligent diagnosis method for common mental illnesses according to claim 1, characterized in that: The multi-agent mental illness diagnostic framework includes: Angel.R model agent: During the diagnosis process, it does not refer to historical case data in the similar patient database, but mainly relies on granular symptom analysis results, user-entered structured medical records, and simplified scale data for reasoning; Angel.D model agent: uses similar historical cases retrieved from a similar patient database as contextual information for reference, but does not conduct in-depth analysis to assist in diagnosis; Angel.C model agent: performs detailed comparison and analysis of similar historical cases retrieved from a database of similar patients, extracting insights from historical experience.

9. The multi-agent-based intelligent diagnosis method for common mental illnesses according to claim 1, characterized in that: After the multi-agent diagnostic debate, the following is also included: The final diagnosis result and detailed reasons supporting the judgment are output. The output results include: clear diagnostic conclusion, detailed reasoning process and supporting evidence.

10. A multi-agent based intelligent diagnosis system for common mental illnesses, characterized by: include: A similar patient knowledge base construction module is used to extract the diagnostic criteria and symptom characteristics of common mental disorders, and to construct a similar patient knowledge base after verification, evaluation, and expert calibration; The clinical data collation and analysis module is used to obtain clinical data from the hospital information system, structure the clinical data into large medical records, simplify the clinical scales, and analyze the scale scores to form a similar patient database; The multi-agent diagnosis module is used to build a multi-agent mental illness diagnosis framework, embed the similar patient knowledge base and similar patient database into the multi-agent mental illness diagnosis framework, increase user input, and conduct multi-agent diagnostic debate.

Images