A method and system for diagnosing and treating adolescent depression based on large model cooperation

Abstract

The application provides a kind of big model cooperation-based adolescent depression intelligent diagnosis and treatment method and system, the method comprises: constructing remote sensing human settlement microenvironment and growth soft environment dataset, pre-processing and feature extraction are carried out on the dataset, space-time matching and representation are carried out on environment and subject individual brain-behavior clinical data, individualized intervention cohort data and feature representation matched with environment are collected;Based on the multidimensional data obtained, a knowledge graph based on large model is established;Based on the established knowledge graph, environment-brain-behavior enhanced intelligent diagnosis and intervention are carried out.The application is based on the knowledge organization system of auxiliary diagnosis, pathogenesis evaluation and individualized intervention of environment-brain-behavior, constructs environment-brain-behavior multidimensional knowledge graph for adolescent depression diagnosis, based on the big model cooperative reasoning mechanism enhanced by environment-brain-behavior multidimensional knowledge, provides intelligent and precise method system for adolescent depression diagnosis and treatment, greatly improves the level of diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the fields of knowledge graphs and large model technologies, and more specifically, to an intelligent diagnosis and treatment method and system for adolescent depression based on large model collaboration. Background Technology

[0002] The high incidence and rapid rise in adolescent depression have become a prominent public health issue. Faced with the reality that a shortage of psychiatric personnel cannot meet the large-scale demand for diagnosis and treatment, intelligent diagnosis and treatment, and improved efficiency, are wise solutions. Artificial intelligence technologies, primarily knowledge graphs and large-scale modeling, are rapidly integrating with clinical psychiatric diagnosis and treatment, forming new service models.

[0003] The pathogenesis of adolescent depression remains unclear, primarily due to insufficient exploration of the "environment-brain network-behavior" relationship. From the perspective of the impact of environmental factors on adolescents' daily lives, these factors can be categorized into the microenvironment of human habitation (such as atmospheric environment, geographical environment, and living environment) and the "soft" environment of growth. Preliminary studies have revealed that atmospheric pollutants are risk factors for depression. Geographical environments such as forest vegetation and sunlight may trigger depressive symptoms. The living environment, including pedestrian and vehicular density and noise, has also been preliminarily confirmed to be related to the risk of depression. The "soft" environment of growth, such as parenting styles, childhood trauma experiences, and support from relatives and peers, plays a significant role in the occurrence and development of adolescent depression. These environmental factors have been listed as important targets for adolescent depression intervention. There is a complex and interactive relationship between "environment-brain-behavior." Environmental factors can induce depressive behaviors by affecting the function and structure of brain networks; abnormalities in brain networks may make individuals more sensitive to environmental stress, further exacerbating depressive symptoms; and depressive behaviors (such as school aversion and social withdrawal) may further alter an individual's environmental exposure patterns (such as reducing outdoor activities), creating a vicious cycle. The aforementioned complex relationships collectively constitute the "environment-brain-behavior" theoretical system for adolescent depression. By fully exploring the relationships among these three elements based on artificial intelligence technologies (knowledge graphs, large-scale modeling), we can provide a strong scientific basis for constructing multi-dimensional knowledge graphs to assist in precise diagnosis and treatment, ultimately clarifying the pathogenesis and intervention mechanisms, and assisting in precise individualized diagnosis and treatment.

[0004] Most research on intelligent diagnosis and treatment in psychiatry uses knowledge graphs or large-scale models alone to assist in diagnosis and treatment. Using large-scale models alone lacks definitive logical knowledge support and can lead to serious hallucination problems; while knowledge graphs have strong professional knowledge and logical connections, effectively enhancing the analytical and reasoning capabilities of large-scale models in specific domains. Using knowledge graphs alone without combining them with large-scale models results in incomplete knowledge coverage, insufficient generalization ability, limited reasoning ability, poor scalability, and limited application scenarios. Combining large-scale models can effectively compensate for these shortcomings and improve overall performance and application effectiveness. Sarabadani combined the application of knowledge graphs and large-scale models in neuroscience to construct and evaluate the PKG-LLM framework, aiming to improve methods for assessing and treating depression. Yin proposed a neurosymbolic multi-agent framework that synergizes knowledge graphs and large-scale models, providing psychiatrists with professional diagnostic results and treatment recommendations. While preliminary studies have combined knowledge graphs and large-scale models for intelligent diagnosis, these studies have not focused on the relationship between the environment (human microenvironment, growth "soft" environment) and the brain and behavior, nor have they comprehensively integrated multi-dimensional knowledge of "environment-brain-behavior" with large-scale models to assist in the individualized diagnosis and treatment of adolescent depression. Therefore, this invention will design an intelligent diagnosis and treatment method and system for adolescent depression that integrates "environment-brain-behavior" knowledge with large-scale models, aiming to more accurately and intelligently assist in the diagnosis and treatment of adolescent depression. Summary of the Invention

[0005] This invention addresses the technical problems existing in the prior art by providing an intelligent diagnosis and treatment method and system for adolescent depression based on large-scale model collaboration. This invention combines a multi-dimensional knowledge graph of "environment-brain-behavior" with large-scale model collaboration technology to conduct research on intelligent diagnosis and treatment of adolescent depression, which can provide an intelligent and precise methodological system for the diagnosis and treatment of adolescent depression, and significantly improve the level of diagnosis and treatment.

[0006] According to a first aspect of the present invention, a method for intelligent diagnosis and treatment of adolescent depression based on large-scale model collaboration is provided, comprising:

[0007] We constructed a remote sensing dataset of human living microenvironment and growth soft environment, preprocessed and extracted features from the dataset, spatiotemporally matched and represented the environment with the individual subject's "brain-behavior" clinical data, and collected individualized intervention cohort data and feature representations of the environment matching.

[0008] A knowledge graph based on a large model is built based on the acquired multi-dimensional data;

[0009] Intelligent diagnosis and intervention based on the established knowledge graph to enhance the environment, brain, and behavior.

[0010] Based on the above technical solution, the present invention can also be improved as follows.

[0011] Preferably, the human settlement microenvironment includes: human settlement microenvironment-atmospheric environmental factors, human settlement microenvironment-geographical environmental factors, and human settlement microenvironment-living environmental factors; the growth soft environment includes the characteristic values ​​of the subject in the growth soft environment.

[0012] Preferably, the characteristic values ​​of the growth environment are assessed and quantified using the Childhood Abuse Questionnaire, the Parenting Style Evaluation Scale, and the Social Support Rating Scale. The total score of each scale is the characteristic value of the subject in the growth environment.

[0013] Preferably, the preprocessing and feature extraction of human settlement microenvironment-atmospheric environmental factors includes:

[0014] Based on ultraviolet band data from the Luojia series satellites, the concentration of trace gas columns was retrieved using differential optical absorption spectroscopy algorithm, thereby achieving near-ground concentration retrieval.

[0015] Preferably, the preprocessing and feature extraction of human settlement microenvironment-geographical environment factors include:

[0016] Light pollution information extraction involves acquiring remote sensing data of daylight and nighttime light, and calculating the light pollution index to characterize pollution intensity.

[0017] Forest vegetation information is extracted, vegetation indices are calculated, and forest vegetation intensity is characterized.

[0018] Preferably, the preprocessing and feature extraction of human settlement microenvironment-living environment factors includes:

[0019] Population density information extraction: extracting the temporal resolution of population density data from nighttime light remote sensing images;

[0020] Traffic flow density information extraction: Traffic flow information is extracted from traffic system data and road network density information obtained from satellites.

[0021] Preferably, the spatiotemporal matching and characterization of the environment with the individual subject's "brain-behavior" clinical data includes:

[0022] The study will perform spatiotemporal matching at the street level and below with the subjects' usual residences for several years before the onset of the disease, and track the correlation between the human settlement microenvironment factors and individual clinical data of adolescents with depression, including the soft environment of growth, brain networks, behavioral data and clinical data.

[0023] Complete the spatiotemporal window matching of environmental factors and disease cohorts to form a spatiotemporally matched database of "environment-patient residence-patient clinical data".

[0024] Preferably, the multi-level knowledge organization framework includes: a narrative layer, an event layer, a concept layer, and an attribute layer; the narrative layer describes knowledge related to the diagnosis and treatment of adolescent depression; the event layer uses a contextual graph to describe daily life events and diagnosis-related events of the disease, and can represent the relationships between events;

[0025] The conceptual layer models the common medical conceptual relationships of the disease; the attribute layer uses data attributes to characterize the diagnostic and treatment entities of adolescent depression.

[0026] Preferably, the intelligent diagnosis and intervention based on the established knowledge graph for environment-brain-behavior enhancement includes:

[0027] Get preliminary diagnostic advice;

[0028] By using a multi-round suggestion generation method, the system verifies whether there are hallucination issues in the preliminary diagnostic suggestions. If so, the suggestions are eliminated. The final diagnostic suggestions are then generated based on the remaining suggestions.

[0029] The study analyzes the influence of various factors in diagnostic recommendations and uses the Shapley value to identify potential objective biomarkers for diagnosis. Based on these potential objective biomarkers, it identifies biomarkers related to disease and the environment, feeds them into a large-scale model, and allows the model to generate personalized intervention recommendations that match the environment based on these environmental biomarkers.

[0030] By employing a multi-round suggestion generation method, the illusion problem of intervention suggestions that are matched to the environment is reduced.

[0031] According to a second aspect of the present invention, a smart diagnostic and treatment system for adolescent depression based on large-scale model collaboration is provided, comprising:

[0032] The data acquisition module is used to construct a remote sensing dataset of human living microenvironment and growth soft environment, preprocess the dataset and extract features, perform spatiotemporal matching and characterization of the environment and individual subject "brain-behavior" clinical data, and collect individualized intervention cohort data and feature representations of environment matching.

[0033] The knowledge graph construction module is used to build a knowledge graph based on a large model based on the acquired multi-dimensional data.

[0034] The intelligent diagnosis and intervention module is used for intelligent diagnosis and intervention based on the established knowledge graph to enhance the environment, brain, and behavior.

[0035] The technical effects and advantages of this invention are as follows:

[0036] This invention provides an intelligent diagnosis and treatment method and system for adolescent depression based on large-scale model collaboration. By establishing a knowledge organization system based on "environment-brain-behavior" for auxiliary diagnosis, pathogenesis assessment and individualized intervention, it constructs a multi-dimensional knowledge graph of "environment-brain-behavior" for the accurate diagnosis of adolescent depression. Based on the large-scale model collaborative reasoning mechanism enhanced by the multi-dimensional knowledge of "environment-brain-behavior", it realizes the mining of objective diagnostic markers, pathogenesis research and individualized intervention guidance that matches the environment for adolescent depression patients. Attached Figure Description

[0037] Figure 1 A flowchart illustrating an intelligent diagnosis and treatment method for adolescent depression based on large-model collaboration, provided as an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of knowledge extraction and multi-relation modeling based on a large model provided in an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram illustrating personalized intervention guidance based on a multimodal diagnostic and therapeutic intelligent agent, provided in an embodiment of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Adolescent depression has a high incidence rate and significant harm, becoming a prominent public health issue. Its diagnosis and treatment rely heavily on physicians' subjective experience and lack objective biomarkers. The pathogenesis and intervention mechanisms involve complex interactions among multiple factors, including environment, brain networks, and behavior, making individualized and precise diagnosis and treatment difficult.

[0042] Understandably, given the deficiencies in the background technology, this invention proposes an intelligent diagnosis and treatment method for adolescent depression based on large-scale model collaboration, specifically as follows: Figure 1 As shown, it includes:

[0043] Step S1. Obtain multi-dimensional data on environment, brain, and behavior; including:

[0044] 1.1. Construct a remote sensing dataset of human settlement microenvironment and growth "soft" environment, and perform preprocessing and feature representation.

[0045] In this embodiment, high spatial resolution environmental factor information extraction is achieved based on observation data from the Luojia series satellites, represented by the "People's Health" satellite. Remote sensing data acquisition and processing of different environmental factors are implemented to address their unique attributes; a remote sensing human settlement microenvironment dataset is constructed, and preprocessing and feature representation are performed. The subject's developmental "soft environment" is primarily assessed and quantified using the Childhood Trauma Questionnaire (CTQ), the Inventory for Measurement of Parental Rearing Pattern, and the Social Support Rating Scale. The total score of each scale represents the subject's characteristic value in this developmental "soft" environment.

[0046] The aforementioned human settlement microenvironment and growth "soft" environment include: human settlement microenvironment - atmospheric environmental factors, human settlement microenvironment - geographical environmental factors, human settlement microenvironment - living environment factors, and growth "soft" environment characteristics.

[0047] Specifically, the preprocessing and feature extraction of human settlement microenvironment-atmospheric environmental factors include:

[0048] Based on ultraviolet band data from the Luojia series satellites, including "People's Health", the differential optical absorption spectroscopy (DOAS) algorithm was used to invert the concentration of trace gas columns, and then the near-ground concentration was inverted.

[0049] Slanted column concentration (SCD) inversion based on spectral fitting. A relationship between entrance pupil radiance and atmospheric species (trace gases) is constructed based on Beer-Lambert's law. In the DOAS method, the contributions of molecular (Rayleigh) scattering and aerosol scattering to solar radiation attenuation are separated from high-frequency signals generated due to molecular absorption. The reflectance spectrum is divided into spectrally smoothed portions. (Broadband absorption and scattering) and the difference component. After removing the smoothing effect, the trace gas column concentration (SCD) is obtained based on the spectral fitting of the trace gas and the entrance pupil spectrum.

[0050]

[0051] in --Solar irradiance at the entrance pupil, --Extraterrestrial solar irradiance -- is the density of the column of unknown species. --Absorption cross-section of absorbing species, --wavelength polynomial, --depending on Photon noise.

[0052] The second step is to calculate the Air Quality Factor (AMF) and convert the oblique column to a vertical column concentration (VCD). AMF represents the ratio of the effective light path through the atmosphere to the vertical light path. AMF is calculated based on the radiative transfer model (SCIATRAN) and prior trace gas profiles. For scattering weights, This is the shape factor.

[0053]

[0054] The third step involves using machine learning methods and with the assistance of geographic and meteorological data to construct the relationship between the concentration of trace gases from satellite columns and the concentration observed at near-surface stations, thereby achieving the inversion of near-surface trace gas concentrations.

[0055]

[0056] Human settlement microenvironment - Geographical environment factor preprocessing and feature extraction includes:

[0057] Geographical environmental factors include light pollution, forest vegetation, and park water systems. Specific satellites and wavelengths are required for information extraction depending on the type of land cover or the object of observation.

[0058] Light pollution information extraction: Based on the observation constellation consisting of three satellites, "Luojia-1" 01, "Qimingxing" and "Luojia-3" 02, remote sensing observation data of sunlight and night light are acquired, and the light pollution index is calculated to characterize the pollution intensity.

[0059]

[0060] in, The ratio of the radiance of the nighttime light remote sensing image of the area to be evaluated to that of the reference area (a remote area with lower light pollution levels).

[0061] Forest vegetation information extraction: Based on the red and near-infrared bands of the "People's Health" satellite, vegetation indices are calculated to characterize forest vegetation intensity.

[0062]

[0063] in, The Normalized Difference Vegetation Index (NDVI) is one of the most commonly used vegetation indices in the field of remote sensing. This is the near-infrared radiation value. This represents the radiation value in the red light band.

[0064] Human settlement microenvironment - Preprocessing and feature extraction of living environment factors include:

[0065] Environmental factors include information such as pedestrian density and vehicle density. It is necessary to establish a correlation between satellite data and existing statistical results to extract information on these environmental factors.

[0066] Population density information extraction: The existing population density data (Worldpop population density data) has a time resolution of one year, which cannot achieve fine information extraction. The time resolution of the population density data can be improved by using nighttime light remote sensing imagery.

[0067]

[0068] in, Original population density, Daily population density For daily nighttime light brightness, It refers to the number of days in a year.

[0069] Traffic flow density information extraction: Traffic flow information is extracted by combining traffic system data with road network density information acquired from satellites. Road network density can be calculated based on road shape obtained from satellite hyperspectral imagery.

[0070]

[0071] in, The length of each road in the grid, The width of each road in the grid. The area is the grid area.

[0072] The extraction of characteristics of the growth "soft" environment includes:

[0073] The subject's "soft environment" of growth was mainly assessed and quantified using the Childhood Trauma Questionnaire (CTQ), the Inventory for Measurement of Parental Rearing Pattern, and the Social Support Rating Scale (SSRS). The total score of each scale is the subject's characteristic value in this "soft" environment of growth.

[0074] The scoring criteria for the Childhood Abuse Questionnaire are as follows: (1) Sexual abuse: a total score ≥8 indicates sexual abuse, and <8 indicates no sexual abuse; (2) Emotional abuse: a total score ≥13 indicates emotional abuse, and <13 indicates no emotional abuse; (3) Physical abuse (9, 11, 12, 15, 17): a total score ≥10 indicates physical abuse, and <10 indicates no physical abuse; (4) Physical neglect (1, 2, 4, 6, 26): a total score ≥10 indicates physical neglect, and <10 indicates no physical neglect; (5) Emotional neglect (5, 7, 13, 19, 28): a total score ≥15 indicates emotional neglect, and <15 indicates no emotional neglect. The total score of each dimension, including sexual abuse, physical abuse, emotional abuse, physical neglect, and emotional neglect, is the characteristic value of the subject's childhood abuse.

[0075] Parenting Style Evaluation Scale: The standard deviations and means of both parents on each dimension are shown in Tables 1 and 2. The total scores of the subjects on the dimensions of father-emotional warmth / understanding, father-punishment / strictness, father-over-interference, father-over-protection, father-rejection / denial, mother-emotional warmth / understanding, mother-punishment / strictness, mother-over-interference / over-protection, and mother-rejection / denial represent the characteristic values ​​of the subjects' parenting styles.

[0076] Table 1. Standard deviation of the father's side in each dimension

[0077]

[0078] Table 2. Standard deviation of the mother in each dimension

[0079]

[0080] The Social Support Rating Scale has a total score of 40 points. Scoring criteria: a total score < 20 indicates "low social support"; 20 ≤ total score < 30 indicates "moderate social support"; and 30 ≤ total score ≤ 40 indicates "satisfactory social support". The total score represents the subject's characteristic value in terms of social support.

[0081] 1.2. Spatiotemporal matching and representation of environmental and individual subject "brain-behavior" clinical data:

[0082] Precise spatiotemporal matching of remote-sensed "human settlement microenvironment" factors with the subjects' usual residences in the five years prior to the onset of illness was performed at a street-level and lower resolution (approximately 10 to 50 meters) to track the correlation between human settlement microenvironment (atmosphere, geography, and living environment) factors and individual-level clinical data of adolescents with depression, including the soft environment of growth, brain networks, and behavioral data (mainly clinical scales and examination results). Spatiotemporal window matching of environmental factors and the disease cohort was also completed, forming a spatiotemporally matched "environment-patient's usual residence-patient clinical data" database, providing data support for the analysis of the relationship between environmental factors and disease.

[0083] Based on the remote sensing human settlement microenvironment database, human settlement microenvironment factors are ultimately processed into gridded daily-level images. The human settlement microenvironment factor data undergoes spatial resampling to form grid data with uniform spatial resolution, meeting the requirements of street-level analysis. Subsequently, based on the subject's permanent residence database, the latitude and longitude information of the subject's residence is extracted using map tool APIs to complete the spatial matching of remote sensing human settlement microenvironment factors and health data. Next, the spatiotemporal window matching of remote sensing human settlement microenvironment factors and subject cohort data is completed. Through the diagnostic information of the disease cohort, a dynamic disease monitoring window is constructed. The environmental factor images within the window time are temporally reconstructed with the sample spatial matching data, ultimately forming a spatiotemporally matched "environment-patient permanent residence-patient clinical data" database, providing data support for the analysis of the relationship between remote sensing human settlement microenvironment factors and diseases.

[0084] 1.3. Data collection and feature representation of individualized intervention cohorts for environment matching.

[0085] During hospitalization, patients were given a comprehensive treatment plan using commonly used antidepressants. Scales were used to assess changes in clinical symptoms and cognitive function after treatment to clarify the changes in efficacy. Throughout the entire course of treatment—acute, consolidation, and maintenance phases—individualized intervention guidance tailored to the environment was provided. Patients' temporal changes in their condition were followed up, and a longitudinal cohort predicting the full-cycle intervention effect was further constructed. The scale scores at each treatment stage represent the temporal cohort data characteristics related to the intervention effect.

[0086] Patients received an 8-week course of conventional clinical treatment including medication, physical therapy, and psychotherapy. Building upon this traditional approach, personalized intervention plans were designed to match the individual patient's local environment and family "soft" environment. Professional guidance tailored to each patient's needs was provided (e.g., analysis revealed that parental arguments and high levels of noise in the living environment exacerbated depressive behavior in certain adolescent patients, leading to suggestions such as creating a harmonious family environment and reducing noise pollution). Throughout the acute phase (8 weeks), consolidation phase (3 months), and maintenance phase (9 months), individualized intervention guidance was provided, and the patient's temporal changes in condition were monitored. Furthermore, longitudinal cohort data predicting the intervention's effectiveness throughout the entire treatment course were collected. Patient scores on scales at each treatment stage represent the temporal cohort characteristics related to the intervention's effectiveness. Through empirical research, we can identify objective biomarkers that can predict the effectiveness of patient interventions (changes in scale scores at each treatment stage), such as brain networks, facial expressions, and speech.

[0087] Step S2. Establish a multi-dimensional, multi-modal knowledge graph based on a large model using multi-dimensional data;

[0088] Adolescent depression, as a complex mental illness, involves multi-dimensional medical knowledge encompassing the "environment-brain-behavior" framework in its diagnosis and treatment. Traditional knowledge extraction methods often face challenges such as entity ambiguity, relational ambiguity, and difficulties in fusing multi-dimensional and multimodal data when dealing with this complex medical knowledge. Therefore, this invention, through large-model-guided knowledge extraction and multi-relational modeling, constructs a precise and scalable multi-dimensional multimodal psychological knowledge graph (MM-PsyKG) related to the diagnosis and treatment of adolescent depression, encompassing the "environment-brain-behavior" framework.

[0089] To more effectively organize medical knowledge about adolescent depression, embodiments of this invention construct a multi-dimensional, multi-modal knowledge graph, such as... Figure 2As shown, the multi-dimensional, multimodal knowledge graph involves a narrative layer, an event layer, a concept layer, and an attribute layer. The attribute layer uses data attributes to characterize entities related to the diagnosis and treatment of adolescent depression. The concept layer models common medical conceptual relationships related to the disease. The event layer uses a contextual graph to describe daily life events and treatment-related events related to the disease, and can represent the relationships between events. The narrative layer is a novel layer that describes knowledge related to the diagnosis and treatment of adolescent depression. The purpose of introducing this layer is to organize various types of psychiatric knowledge under a goal-driven approach, and to implement knowledge representation modeling based on the causal relationships, traceability, uncertainty, and subjectivity of adolescent depression diagnosis and treatment information, thereby ensuring that relationships are effectively established and more comprehensively supporting downstream tasks. The multi-layered knowledge organization framework (i.e., a four-layer framework of narrative, event, concept, and attribute layers) mentioned in this paragraph is constructed using the following methods: In the concept and attribute layers, the Clinical Medical Knowledge Ontology (HealthOnto) and the Evidence-Based Medical Literature Ontology (MDO) are mainly used to represent basic knowledge in clinical practice and literature. In the event layer, the core elements of HealthOnto and MDO are integrated, with treatment events as the center.

[0090] Step S3. Based on the established knowledge graph, conduct intelligent diagnosis and intervention for environment-brain-behavior enhancement.

[0091] This invention combines a multi-dimensional "environment-brain-behavior" knowledge graph with existing large-scale models. The aim is to enhance the large-scale models' ability to analyze professional medical knowledge, intelligently mine objective biomarkers for disease diagnosis, and alleviate the hallucination problem that may occur during logical reasoning in large-scale models, thereby constructing an auxiliary diagnostic and treatment intelligent agent based on a large-scale model. Specifically, as follows... Figure 3 As shown, this section mainly includes four aspects: a knowledge-enhanced architecture for assistive diagnosis and treatment; correction of large model illusions; mining of objective biomarkers for diagnosis; and intervention guidance based on environment matching of the diagnostic and treatment agent. This invention corrects large model illusions, mines potential objective biomarkers for assistive diagnosis, and achieves personalized intervention guidance based on agent-based environment matching by constructing a knowledge-enhanced architecture for assistive diagnosis and treatment.

[0092] The intelligent diagnosis and intervention based on the established knowledge graph that enhances the environment, brain, and behavior includes:

[0093] First, we receive a "preliminary diagnostic suggestion";

[0094] By using a multi-round suggestion generation method, the system verifies whether there are hallucination issues in the "preliminary diagnostic suggestions". If so, they are filtered out. Then, based on the remaining suggestions, the final diagnostic suggestions are generated.

[0095] The study analyzes the influence of various factors in diagnostic recommendations, uses Shapley values ​​to mine potential objective biomarkers for diagnosis, and then mines disease-related and environment-related biomarkers based on these potential objective biomarkers. These biomarkers are then fed into a large model, which generates personalized intervention recommendations that match the environment based on these environmental biomarkers. Finally, through multiple rounds of recommendation generation, the illusion of "environmentally matched intervention recommendations" is reduced.

[0096] Specifically, Figure 3 The diagram illustrates personalized intervention guidance based on a multimodal diagnostic agent provided in this embodiment of the invention. As shown, the knowledge-enhanced auxiliary diagnostic agent architecture constructed in this embodiment mainly consists of a retrieval routing agent, a retrieval tool library, and a generative large model. The retrieval routing agent is handled by a large model (such as Deepseek or ChatGPT), primarily used to determine the optimal tool in the retrieval tool library for processing. The retrieval tool library integrates the MM-PsyKG knowledge base, diagnostic guidelines, and web search tools constructed in Part II of this invention. When a user asks a question, the large model first determines whether the question falls within a professional scope. If it is a professional question, the encoding model uses the large model to perform unified feature encoding on the multimodal data (such as audio / video and brain network data). The MM-PsyKG knowledge base retrieves the patient's past diagnostic records and the most relevant written theoretical materials related to the input information. This information is input as background knowledge prompts into the large model to assist in generating a more accurate answer. When handling routine questions, the retrieval routing agent uses web search engine tools to enhance the retrieval effect, optimizes the user's question, and thus assists in generating preliminary diagnostic suggestions.

[0097] Secondly, perform large-scale model illusion correction to correct erroneous outputs and obtain correct outputs.

[0098] The knowledge-enhanced assisted diagnostic agent architecture outputs feedback, but it suffers from erroneous outputs, known as illusions (i.e., incorrect outputs). To effectively address the illusion problem when large models generate content and significantly improve the accuracy of the reasoning process, this invention investigates a self-consistency detection mechanism based on agent retrieval enhancement. This mechanism employs an integrated response self-consistency detection method, which, through multiple generation and feedback loops, prompts the model to continuously self-verify during the generation process, ensuring the stability and accuracy of the output results.

[0099] During the multi-round generation and validation process, the large model simultaneously analyzes the subjective and objective manifestations of patients with various possible conclusions. This process helps construct a potential candidate set of objective diagnostic biomarkers, providing a more accurate and scientific reference for medical diagnosis. In specific implementation, before generating the final diagnostic suggestion, the system performs multiple rounds of independent generation based on the patient's historical data, background knowledge, and current input, employing specific randomization techniques to avoid biases in the generation process affecting the results. For multiple response results, the confidence level of a single inference is calculated by analyzing the similarity between these independently generated results and performing clustering. The results after multiple rounds of independent generation and integration are finally submitted to doctors or clinical decision support systems as a reference for auxiliary diagnosis. If the results from multiple rounds of generation are highly consistent, it indicates that the model has a strong confidence in the diagnostic conclusion and can provide more accurate diagnostic support for clinicians. Conversely, if the results are inconsistent, the system proactively indicates that the reliability of the generated results is low and suggests that they be manually annotated and added to the knowledge base for dynamic updates.

[0100] Next, objective biomarker mining for diagnosis is performed. After correct output, disease-related and environment-related biomarkers are mined from potential objective biomarkers, resulting in potential biomarkers. These potential biomarkers are then fed back into the large-scale model to form final diagnostic recommendations. By analyzing the Shapley values ​​of the large-scale model's diagnostic results and comparing them with biomarker options in the candidate set of objective diagnostic biomarkers summarized by the large-scale model, clinically verifiable objective diagnostic biomarkers are identified, thereby further improving the accuracy of clinical diagnosis.

[0101] Finally, intervention guidance is based on environment matching of the diagnostic agent; after being fed to the large model, the large model provides suggestions for intervention guidance. When generating suggestions, it jumps to the large model's hallucination correction, then retrieves the hallucination problem, and finally generates an intervention guidance suggestion.

[0102] Based on the aforementioned auxiliary diagnostic and treatment intelligent agents and pathogenesis narrative knowledge, embodiments of the present invention will construct personalized intervention guidance suggestions that are matched with the environment based on multimodal diagnostic and treatment intelligent agents.

[0103] In summary, this invention, based on an auxiliary diagnostic agent and a knowledge base of disease pathogenesis narratives, constructs an innovative multimodal diagnostic agent-based personalized intervention guidance method. The implementation of this method first relies on the auxiliary diagnostic agent to accurately predict the intervention efficacy. After prediction, these predictions are transformed into core prompts to construct intervention method queries, which are then input into the agent. The agent then performs an in-depth search in the disease pathogenesis narrative knowledge base to find environmental mechanisms that match the prediction results. Once a corresponding mechanism is found, the agent can automatically generate a series of prompts related to environmental intervention. These prompts are then input again into the multimodal diagnostic agent to generate personalized intervention guidance plans. In this way, this project will provide patients with personalized intervention plans tailored to their specific circumstances and individual needs, matching their living micro-environment and family "soft" environment, and providing professional guidance suggestions that meet their specific requirements. Through empirical research, it will explore multimodal objective markers such as brain networks, facial expressions, and speech to predict and evaluate the effects of individualized interventions.

[0104] According to a second aspect of the present invention, a large-scale collaborative intelligent diagnosis and treatment system for adolescent depression is provided. This system integrates the collection and extraction of multi-dimensional, multi-modal data on adolescent depression, as well as diagnosis and efficacy prediction, enabling precise diagnosis and treatment of adolescent depression based on an "environment-brain-behavior" model. The system includes:

[0105] The data acquisition module is used to construct a remote sensing dataset of human living microenvironment and growth soft environment, preprocess the dataset and extract features, perform spatiotemporal matching and characterization of the environment and individual subject "brain-behavior" clinical data, and collect individualized intervention cohort data and feature representations of environment matching.

[0106] The knowledge graph construction module is used to build a knowledge graph based on a large model based on the acquired multi-dimensional data.

[0107] The intelligent diagnosis and intervention module is used for intelligent diagnosis and intervention based on the established knowledge graph to enhance the environment, brain, and behavior.

[0108] It is understood that the intelligent diagnosis and treatment system for adolescent depression based on large model collaboration provided by the present invention corresponds to the intelligent diagnosis and treatment method for adolescent depression based on large model collaboration provided in the foregoing embodiments. The relevant technical features of the intelligent diagnosis and treatment system for adolescent depression based on large model collaboration can be referred to the relevant technical features of the intelligent diagnosis and treatment method for adolescent depression based on large model collaboration, and will not be repeated here.

[0109] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A large model cooperation-based intelligent diagnosis and treatment recommendation method for adolescent depression, characterized in that, The method comprises: Remote sensing of human microenvironment and growth soft environment dataset is constructed, the dataset is preprocessed and feature extracted, the environment and the individual "brain-behavior" clinical data of the subject are spatio-temporally matched and characterized, and the individualized intervention cohort data and feature representation matched with the environment are collected; the human microenvironment comprises: human microenvironment-atmospheric environment factor, human microenvironment-geographical environment factor and human microenvironment-living environment factor; the growth soft environment comprises the characteristic value of the subject in the growth soft environment; wherein, The preprocessing and feature extraction of the human microenvironment-atmospheric environment factor comprises: based on the ultraviolet band data of the Luoja series satellite, the trace gas column concentration is inversed by using the differential optical absorption spectrum algorithm, and then the near-surface concentration is inversed; The preprocessing and feature extraction of the human microenvironment-geographical environment factor comprises: light pollution information extraction, obtaining sunlight and night light remote sensing observation data, and calculating the light pollution index to represent the pollution intensity; Forest green plant information extraction, calculation of vegetation index, and representation of forest green plant intensity; The preprocessing and feature extraction of the human microenvironment-living environment factor comprises: population density information extraction, time resolution of population density data extracted by night light remote sensing image; Traffic flow density information extraction, traffic flow information extracted by traffic system data and satellite obtained road network density information; Based on the obtained multi-dimensional data, a knowledge graph based on a large model is established; Based on the established knowledge graph, an environment-brain-behavior enhanced intelligent diagnosis is carried out and intervention suggestions are put forward.

2. The method according to claim 1, wherein, The characteristic value on the growth soft environment is evaluated and quantified by means of the childhood abuse questionnaire, the parent's education method evaluation scale and the social support evaluation scale, and the total score of each scale is the characteristic value of the subject on the growth soft environment.

3. The method according to claim 1, wherein, The spatio-temporal matching and characterization of the environment and the individual "brain-behavior" clinical data of the subject comprises: Spatio-temporal matching of remote sensing human microenvironment factors and the subject's residence for several years before the onset at the block level and below, tracking the human microenvironment factors and the clinical data of the individual level of adolescent depression subjects, including the correlation among the growth soft environment, brain network, and behavior data clinical data; Complete spatio-temporal window matching of environmental factors and patient cohort to form a "environment-patient residence-patient clinical data" database matched in space and time.

4. The method according to claim 1, wherein, The establishment of a knowledge graph based on a large model based on the obtained multi-dimensional data comprises: narrative layer, event layer, concept layer and attribute layer; wherein, the narrative layer is a layer describing the knowledge related to the diagnosis and treatment of adolescent depression; the event layer describes the daily life events and diagnosis and treatment related events of adolescent depression by using the event graph, and can represent the correlation between events; the concept layer models the general medical concept relationship of adolescent depression; and the attribute layer describes the diagnosis and treatment entities of adolescent depression by means of data attributes.

5. The method according to claim 1, wherein, The environment-brain-behavior enhanced intelligent diagnosis based on the established knowledge graph and the intervention suggestions comprise: Obtaining preliminary diagnosis suggestions; Through a multi-round suggestion generation method, it is verified whether there is a hallucination problem in the preliminary diagnosis suggestion, if there is, it is eliminated, and then according to the remaining suggestions, the final diagnosis suggestion is generated. Analyze the influence degree of various influencing factors in the diagnostic suggestion, and mine potential diagnostic objective biomarkers through the Shapley value. According to the potential diagnostic objective biomarkers, environmental markers related to diseases and environments are mined, and the large model is fed to generate individualized intervention suggestions matched with the environment according to the environmental markers. Through the multi-round suggestion generation method, the illusion problem of the intervention suggestions matched with the environment is reduced.

6. A large model cooperation-based intelligent diagnosis and treatment recommendation system for adolescent depression, characterized in that, It comprises: A data acquisition module is used to construct a remote sensing human settlement microenvironment and growth soft environment dataset, pre-process and extract features of the dataset, match and represent the environment and the subject individual "brain-behavior" clinical data in space and time, and collect environment-matched individualized intervention cohort data and feature representation. The human settlement microenvironment includes: human settlement microenvironment-atmospheric environment factor, human settlement microenvironment-geographical environment factor and human settlement microenvironment-living environment factor; the growth soft environment includes the characteristic value of the subject in the growth soft environment; wherein, The human settlement microenvironment-atmospheric environment factor preprocessing and feature extraction includes: based on the Lojia series satellite ultraviolet band data, using the differential optical absorption spectrum algorithm to retrieve the trace gas column concentration, and then realizing the near-surface concentration retrieval; The human settlement microenvironment-geographical environment factor preprocessing and feature extraction includes: light pollution information extraction, obtaining sunlight and night light remote sensing observation data, and calculating the light pollution index to represent the pollution intensity; Forest green plant information extraction, calculation of vegetation index, representation of forest green plant intensity; The human settlement microenvironment-living environment factor preprocessing and feature extraction includes: population density information extraction, extracting the time resolution of population density data through night light remote sensing image; Traffic flow density information extraction, traffic flow information is extracted through traffic system data and satellite obtained road network density information; A knowledge graph construction module is used to establish a knowledge graph based on a large model based on the obtained multi-dimensional data; An intelligent diagnosis and intervention module is used to perform environment-brain-behavior enhanced intelligent diagnosis and propose intervention suggestions based on the established knowledge graph.

Images