Diagnosis and treatment sample construction method and device and auxiliary diagnosis and treatment large model training method and device

Abstract

The embodiment of the invention discloses a diagnosis and treatment sample construction method and device and an auxiliary medical large model training method and device. The construction method of the diagnosis and treatment sample comprises the following steps: firstly, determining a plurality of candidate symptom combinations by utilizing a knowledge graph constructed based on mental disease data; next, for any first candidate symptom combination, reasoning the symptom combination by using a sub-graph corresponding to the diagnosis rule in the knowledge graph to obtain a diagnosis class label; then, reasoning the first candidate symptom combination and the diagnosis class label by using a sub-graph corresponding to a treatment guide in the knowledge graph to obtain a treatment suggestion label; and then, constructing the first candidate symptom combination, the diagnosis class tag and the treatment suggestion tag into a first candidate sample for determining a diagnosis and treatment sample set. Furthermore, based on a high-quality diagnosis and treatment sample set constructed at low cost, training of the auxiliary medical large model can be carried out, so that the trained auxiliary medical large model has good generalization.

Description

Technical Field

[0001] This specification relates to the field of intelligent medical technology in one or more embodiments, and particularly to a method and apparatus for constructing a diagnostic sample, a method and apparatus for training a large-scale auxiliary diagnostic model, a computer-readable storage medium, and a computing device. Background Technology

[0002] With the accelerating pace of modern life and increasing social competition, the incidence of mental health problems such as depression, anxiety, and sleep disorders has risen significantly, becoming a major issue affecting public health. Mental illness not only severely reduces patients' quality of life but also imposes a heavy socioeconomic burden. Therefore, developing efficient and accurate diagnosis and treatment for mental illness has significant social and clinical value.

[0003] In recent years, artificial intelligence (AI) technology has developed rapidly, and its application in the healthcare field has become increasingly widespread, especially in assisted diagnosis and personalized treatment recommendations, demonstrating enormous potential. AI-based assisted diagnosis and treatment technology for mental illnesses can provide doctors with objective decision support by automatically analyzing patients' characteristic data, helping to improve diagnostic consistency, alleviate the problem of uneven distribution of medical resources, and increase diagnostic efficiency.

[0004] However, existing auxiliary diagnostic and treatment technologies for mental illnesses are insufficient to meet the higher requirements of practical applications, such as reducing the cost of obtaining high-quality labeled data, enhancing the interpretability of diagnostic and treatment results, and introducing existing medical knowledge of mental illnesses into the diagnostic and treatment system and keeping up with the updates of knowledge in a timely manner. Summary of the Invention

[0005] This specification describes a method and apparatus for constructing diagnostic and treatment samples, as well as a method and apparatus for training a large-scale auxiliary diagnostic and treatment model, which can meet the higher requirements in the above-mentioned practical applications.

[0006] According to the first aspect, a method for constructing diagnostic and treatment samples is provided. The method includes: using a knowledge graph constructed based on mental illness data to determine several candidate symptom combinations; for any first candidate symptom combination, using a subgraph of the knowledge graph corresponding to a diagnostic rule to infer the symptom combination and obtain a diagnostic label; using a subgraph of the knowledge graph corresponding to a treatment guideline to infer the first candidate symptom combination and the diagnostic label and obtain a treatment suggestion label; and constructing the first candidate symptom combination, the diagnostic label, and the treatment suggestion label as a first candidate sample to determine a diagnostic and treatment sample set.

[0007] In one embodiment, the mental illness data includes multimodal data, specifically including at least one of the following: language modality data, visual modality data, physiological modality data, behavioral modality data, and scale modality data.

[0008] In one embodiment, a knowledge graph constructed based on mental illness data is used to determine several candidate symptom combinations, including: for any first real case data, determining multiple standard symptom terms corresponding to it to form the first candidate symptom combination; the multiple standard symptom terms belong to entities in the knowledge graph.

[0009] In one embodiment, a knowledge graph constructed based on mental illness data is used to determine several candidate symptom combinations, including: extracting symptoms under different modalities from a subgraph of the knowledge graph corresponding to a first diagnostic rule to form the first candidate symptom combination.

[0010] In one embodiment, the plurality of candidate symptom combinations includes a second candidate symptom combination, wherein the symptoms relate to one or more of the following modalities: language, visual, physiological, behavioral, and scale; and / or, the plurality of candidate symptom combinations includes a third candidate symptom combination, which includes historical symptoms and current symptoms.

[0011] In one embodiment, the diagnostic labels include disease diagnosis labels, severity labels, and comorbidity labels.

[0012] In one embodiment, the diagnostic tags and treatment suggestion tags correspond to entities in the knowledge graph.

[0013] In one embodiment, after constructing the first candidate symptom combination, diagnostic label, and treatment suggestion label as a first candidate sample, the method further includes: scoring the first candidate sample using a first large model, and if the score is greater than a predetermined threshold, classifying the first candidate sample into the diagnostic sample set.

[0014] According to the second aspect, a training method for an auxiliary diagnosis and treatment large model is provided. The method includes: transforming a diagnosis and treatment sample set into a fine-tuning sample set containing inference chains using a second large model, wherein the diagnosis and treatment sample set is determined using the method provided in the first aspect; performing supervised fine-tuning of the auxiliary diagnosis and treatment large model using the fine-tuning sample set; and further fine-tuning the auxiliary diagnosis and treatment large model using a reinforcement learning algorithm, wherein the reward score is derived from the performance of the diagnosis and treatment large model on the diagnosis and treatment sample set.

[0015] In one embodiment, the second major model is used to transform the treatment sample set into a fine-tuned sample set containing an inference chain, including: for any first treatment sample in the treatment sample set, the second major model is used to process the first symptom combination in the sample to obtain the corresponding first inference chain and predicted treatment result; if the predicted treatment result successfully matches the treatment label of the first treatment sample, the first symptom combination, the first inference chain and the treatment label are constructed into a first fine-tuned sample.

[0016] In one embodiment, the auxiliary diagnosis and treatment model is further fine-tuned using a reinforcement learning algorithm, including: using the auxiliary diagnosis and treatment model to process symptom combinations sampled from the diagnosis and treatment sample set and outputting a diagnosis and treatment action; determining a reward score by comparing the diagnosis and treatment action with the diagnosis and treatment labels corresponding to the symptom combinations; and adjusting the model parameters in the auxiliary diagnosis and treatment model with the goal of maximizing the expected value of the reward score.

[0017] According to a third aspect, an apparatus for constructing diagnostic and treatment samples is provided. The apparatus includes: a symptom combination determination module configured to determine several candidate symptom combinations using a knowledge graph constructed based on mental illness data; a diagnostic label determination module configured to, for any first candidate symptom combination, infer a diagnostic label using a subgraph of the knowledge graph corresponding to a diagnostic rule; a treatment label determination module configured to, using a subgraph of the knowledge graph corresponding to a treatment guideline, infer a treatment suggestion label from the first candidate symptom combination and the diagnostic label; and a diagnostic and treatment sample set determination module configured to construct a first candidate sample from the first candidate symptom combination, the diagnostic label, and the treatment suggestion label, for determining a diagnostic and treatment sample set.

[0018] According to the fourth aspect, a training device for an auxiliary diagnosis and treatment large-scale model is provided. The device includes: a fine-tuning sample set determination module, configured to transform a diagnosis and treatment sample set into a fine-tuning sample set containing inference chains using a second large-scale model, wherein the diagnosis and treatment sample set is determined using the device provided in the second aspect; a supervised fine-tuning module, configured to perform supervised fine-tuning of the auxiliary diagnosis and treatment large-scale model using the fine-tuning sample set; and a reinforcement learning module, configured to further fine-tune the auxiliary diagnosis and treatment large-scale model using a reinforcement learning algorithm, wherein the reward score is derived from the performance of the diagnosis and treatment large-scale model on the diagnosis and treatment sample set.

[0019] According to a fifth aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method provided in the first or second aspect.

[0020] According to a sixth aspect, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method provided in the first or second aspect.

[0021] In summary, by employing the methods and apparatus disclosed in the embodiments of this specification, a knowledge graph is first constructed based on prior knowledge of mental illnesses (such as multi-source data, multimodal data, etc.), and then labeled data for the diagnosis and treatment of mental illnesses is automatically synthesized based on this knowledge graph. This allows for the acquisition of a large amount of high-quality labeled data at low cost. Furthermore, this labeled data can be used to train a large-scale auxiliary diagnosis and treatment model, thereby effectively improving the generalization ability of the trained large-scale auxiliary diagnosis and treatment model. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the implementation architecture of the improved scheme disclosed in the embodiments of this specification; Figure 2 This is a flowchart illustrating the knowledge graph construction process disclosed in the embodiments of this specification; Figure 3 This is a schematic diagram of the process steps for constructing a diagnostic sample as disclosed in the embodiments of this specification; Figure 4 This is a schematic diagram of the training method for the large-scale auxiliary diagnosis and treatment model disclosed in the embodiments of this specification; Figure 5 This is a functional structural diagram of the device for constructing a diagnostic sample as disclosed in the embodiments of this specification; Figure 6 This is a functional structure diagram of the training device for the large-scale auxiliary diagnosis and treatment model disclosed in the embodiments of this specification. Detailed Implementation

[0024] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0025] As mentioned earlier, existing auxiliary diagnostic technologies for mental illness are insufficient to meet the higher demands of practical applications. Specifically, current technologies generally rely on real case data or artificially constructed datasets, utilizing deep learning models for end-to-end training to classify or predict mental illness states. However, existing technologies have the following shortcomings:

[0026] 1. The cost of obtaining high-quality, large-scale labeled samples is extremely high, while the actual collected data quality varies, with problems such as inconsistent labeling and excessive noise, which limits the generalization ability of the model.

[0027] 2. Deep learning models themselves have a "black box" characteristic, and their decision-making process lacks interpretability, making it difficult for doctors to understand and trust the model's output, thus limiting their application in real clinical settings.

[0028] 3. It is difficult to effectively introduce and integrate prior knowledge in the field (such as clinical guidelines and pathological psychological mechanisms).

[0029] In addition, the applicant observed the following unique characteristics of diagnosing mental illnesses such as depression, anxiety, and sleep disorders, which also pose challenges to auxiliary diagnostic techniques: 1) Symptom overlap: The symptoms of the three diseases overlap, requiring precise differentiation. 2) Individual variability: Different patients exhibit significant differences in symptom presentation, necessitating personalized diagnosis. 3) Dynamic changes: Symptoms change over time, requiring dynamic monitoring. 4) Multifactorial influence: Influenced by physiological, psychological, and social factors. 5) Updated diagnostic criteria: Medical diagnostic criteria are constantly being updated, requiring timely adaptation.

[0030] Based on the above observations and analysis, the applicant proposes an improved auxiliary treatment plan for mental illness (hereinafter referred to as the improved plan). For example... Figure 1 As shown, a knowledge graph is first constructed based on prior knowledge of mental illnesses (such as relevant documents in medical databases). Then, labeled data for mental illness diagnosis and treatment is synthesized based on this knowledge graph, and this labeled data is used to train a large-scale auxiliary diagnosis and treatment model. Thus, 1) constructing a knowledge graph allows for the introduction of prior knowledge and the effective integration of complex relationships between mental illness data. 2) The knowledge graph can be updated according to updates to domain-related data, thereby resynthesizing labeled data to retrain the large-scale auxiliary diagnosis and treatment model, enabling the diagnosis and treatment system to keep pace with knowledge updates. 3) The labeled data is automatically synthesized based on the knowledge graph, resulting in high quality and low cost. 4) The trained large-scale auxiliary diagnosis and treatment model provides highly accurate diagnosis and treatment results, and along with the results, an inference chain (corresponding to one or more inference steps) is generated, making the diagnosis and treatment results verifiable and interpretable.

[0031] Next, for Figure 1 The four sections shown in the image will be introduced separately.

[0032] I. Construction of Knowledge Graph

[0033] See Figure 2 The construction of a knowledge graph can be divided into the following stages, and the entire process can be iterative and cyclical:

[0034] 1. Knowledge definition and pattern design

[0035] It should be understood that this stage generally requires close collaboration between domain experts (doctors, medical experts) and knowledge engineers.

[0036] Specifically, based on the problem the knowledge graph aims to solve, such as assisting doctors in making preliminary diagnoses and providing treatment suggestions, an ontology (schema) is defined. An ontology can be understood as a "mold" for the knowledge graph. Below is a simple example of a disease diagnosis ontology design:

[0037] Entity types: Disease, Symptom, Drug, Treatment, etc.

[0038] Relationship types: Has_Symptom (the disease has symptoms), May_Cause (the disease may cause complications), Used_To_Treat (the drug is used to treat the disease), Belongs_To_Department (the disease belongs to a certain department), Is_Risk_Factor_For (a certain factor is a risk factor for a certain disease).

[0039] Attributes: Disease entities have attributes such as incidence rate, mortality rate, and heritability; symptom entities have attributes such as location, severity, duration, and frequency; drug entities have attributes such as usage and dosage, side effects, and contraindications.

[0040] 2. Knowledge Acquisition and Extraction

[0041] After defining the ontology, the data raw materials are processed (knowledge extraction) and transformed into instances that conform to the ontology specification, and then populated into the framework defined by the ontology (the "mold" of the knowledge graph).

[0042] Regarding the selection of data materials, the applicant considered that the diagnosis of mental illness has multimodal and temporal characteristics, and proposed to integrate multiple core data sources, multimodal data, and standard medical documents as data materials.

[0043] The multimodal characteristics of mental illness diagnosis are reflected in: 1) Language expression characteristics: Depressed patients speak slowly and in a low tone; anxious patients speak quickly and repeat themselves; patients with sleep disorders show significant mood swings when describing their sleep quality. 2) Nonverbal behavioral characteristics: Facial expressions, body movements, eye contact, and other nonverbal cues are of significant diagnostic value. 3) Physiological signal characteristics: Physiological indicators such as heart rate variability, skin conductance, and electroencephalography reflect mood and sleep status. 4) Behavioral pattern characteristics: Abnormal behavioral patterns such as daily activity regularity, changes in social behavior, and sleep-wake cycles. 5) Scale assessment characteristics: Standardized scales such as PHQ-9, GAD-7, and PSQI provide quantitative assessment criteria.

[0044] The temporal characteristics of mental illness diagnosis are reflected in: 1) Symptom evolution patterns: Depressive symptoms usually progress chronically, anxiety symptoms may have acute onset, and sleep disorders have periodic characteristics. 2) Comorbidity: Depression and anxiety often coexist, and sleep disorders may be secondary symptoms of depression or anxiety. 3) Treatment response timing: Different diseases have different response times and patterns to treatment. 4) Relapse prediction timing: Predicting the risk of disease relapse based on historical symptom changes.

[0045] Data materials may include:

[0046] 1) Multiple core data sources: medical literature databases (authoritative databases such as PubMed), clinical guideline databases (diagnostic standards such as DSM-5 and ICD-11), expert experience databases (clinical experience of psychiatrists), patient databases (real case data), and drug information databases (information on drug mechanisms of action and side effects).

[0047] 2) Multimodal data: Language modality data: patient self-reported symptoms, doctor's consultation records, and voice tone feature analysis; Visual modality data: facial expression recognition, body movement analysis, and eye contact patterns; Physiological modality data: heart rate variability, skin conductance, electroencephalogram (EEG), and sleep monitoring data; Behavioral modality data: daily activity records, social behavior patterns, and sleep-wake cycles; Scale modality data: assessment results of standardized scales such as PHQ-9, GAD-7, PSQI, HAMD, and HAMA.

[0048] 3) Standardized Medical Documents: Diagnostic Scale Library: Integrates standardized diagnostic scales related to depression, anxiety, and sleep disorders; Assessment Tool Library: Collects various psychological assessment tools and scoring standards; Clinical Pathway Library: Establishes standardized clinical pathways for diagnosis and treatment; Expert Consensus Library: Integrates diagnostic consensus and treatment recommendations from psychiatric experts; Case Template Library: Constructs standardized medical record templates and diagnostic report formats.

[0049] As described above, through knowledge acquisition and extraction, we can obtain knowledge instances with clear structure and semantics, forming a high-quality knowledge base. This process does not transform the data raw materials into triples (entity-relation-entity or entity-attribute-value) with unified semantic representation, such as (depressive disorder, Has_Symptom, loss of interest), or (**medicine, side effects, bitter taste in mouth).

[0050] 3. Knowledge integration and storage

[0051] Knowledge extracted from different sources often contains a lot of duplication and conflict, requiring cleaning and fusion:

[0052] 1) Entity alignment: Determine whether two entities from different data sources point to the same object in the real world. If so, they need to be merged into the same entity.

[0053] 2) Conflict Resolution: Knowledge for handling contradictions. For example, one source says "Drug A treats disease B," while another source says "Drug A is contraindicated for disease B." This needs to be processed according to rules such as the authority of the data source and timestamps, or it can be resolved by expert judgment.

[0054] 3) Data Storage: The cleaned and merged triplet data is stored in a dedicated graph database. Commonly used graph databases include Neo4j, Nebula Graph, and JanusGraph. They have made significant optimizations for graph traversal and querying.

[0055] 4. Knowledge Verification and Evaluation

[0056] A knowledge graph that has been initially constructed generally cannot be used directly and requires rigorous validation, such as:

[0057] 1) Quantitative evaluation: Calculate indicators such as accuracy and recall.

[0058] 2) Qualitative assessment: Domain experts review the randomly sampled triplets, such as whether the checks (depressive disorder, has_Symptom, depressed mood) are correct and whether the relationship definition is reasonable.

[0059] 5. Refinement and encapsulation of knowledge

[0060] After the "knowledge verification and evaluation" phase, a high-quality, structured medical knowledge graph was initially constructed, which includes a large number of triples. However, a single triple can only answer "what" (e.g., what are the symptoms of depressive disorders). In practical applications, it is hoped that the knowledge graph can also answer "what to do" (e.g., what to do if symptoms such as insomnia and low mood occur? How to treat them?). Therefore, the knowledge needs to be refined, and its core outputs include diagnostic rules and treatment guidelines.

[0061] 1) Diagnostic rules and treatment guidelines: as explicit subgraphs of encapsulated knowledge

[0062] Diagnostic rules and treatment guidelines are essentially explicit subgraphs or rule-based knowledge formed by refining, combining, and encapsulating knowledge scattered across a vast number of triples according to clinical logic. Their relationship to triples is as follows:

[0063] ① Triples are raw materials: Basic triples, such as (depressive disorder, Has_Symptom, loss of interest), are the "atomic facts" that constitute rules / guidelines and are the cornerstone of knowledge graphs.

[0064] ② Rules / Guidelines are finished products: Diagnostic rules and treatment guidelines are "molecular structures" assembled from these "atomic facts" through logical operators (such as "AND", "OR", and "NOT"), probability weights, and clinical context. They enable knowledge graphs to have direct clinical reasoning capabilities.

[0065] 2) Construction of diagnostic rules

[0066] This stage requires the deep involvement of domain experts, aiming to transform implicit clinical thinking patterns into explicit structures.

[0067] ① Rule Extraction and Modeling: Knowledge engineers collaborate with clinical experts to transform authoritative diagnostic criteria from medical literature on mental illnesses, such as the DSM-5 / ICD-11 diagnostic criteria, into a computable structure. This is not just a simple extraction, but a process of remodeling on top of the ontology.

[0068] ② First, create the rule entity: Create an entity node for a diagnostic rule (such as "DSM-5 diagnostic criteria"). Then, establish connections: Connect this rule node with all its necessary conditions (symptoms, signs, test results, etc.) using the rule_premise_include relationship, and define logical operators and weights for these relationships. This forms a clear, queryable, and reasonable subgraph structure. Finally, link the conclusion: Connect this rule node with the conclusion (such as depressive disorder) using the rule_pointer_disease relationship.

[0069] To aid understanding, the following example uses the diagnostic criteria for "major depressive disorder" in the DSM-5 (Diagnostic and Statistical Manual of Mental Disorders, 5th Edition) as an example to illustrate a diagnostic rule. The DSM-5 diagnostic criteria require that, within the same two-week period, at least five of the following nine symptoms must be present, including at least one of depressed mood or loss of interest or pleasure, resulting in impaired social functioning. This rule can be modeled as a series of interrelated triples, forming a complete, computable diagnostic logic subgraph:

[0070] a. Define the rule itself and its overall attributes:

[0071] (DSM-5 Diagnostic Criteria for Depressive Disorders, Yes, Diagnostic Rule) / / Declare this entity as a diagnostic rule

[0072] (DSM-5 Diagnostic Criteria for Depressive Disorders, Rule Name: "DSM-5 Diagnostic Criteria for Major Depressive Episode")

[0073] (DSM-5 diagnostic criteria for depressive disorders, minimum number of symptoms required, 5) / / At least 5 symptoms must be met.

[0074] (DSM-5 diagnostic criteria for depressive disorders, duration requirement, "within the same 2-week period") / / Time range attribute

[0075] (DSM-5 diagnostic criteria for depressive disorders, rule pointing to major depressive disorder) / / The rule concludes that the patient has this disorder.

[0076] b. Clearly define the core logic of the rule ("and must include at least one of the following"):

[0077] (DSM-5 diagnostic criteria for depressive disorders, with a prerequisite of depressed mood) / / Core symptom 1

[0078] (DSM-5 diagnostic criteria for depressive disorders, with a prerequisite including loss of interest or pleasure) / / Core symptom 2

[0079] (The core logic of the above two points, the operator, 'OR') / / Meeting either of these two core symptoms is sufficient.

[0080] c. List all other symptom prerequisites included in the rule:

[0081] (DSM-5 diagnostic criteria for depressive disorders, with a prerequisite including significant changes in weight or appetite) / / Symptom 3

[0082] (DSM-5 diagnostic criteria for depressive disorders, with insomnia or excessive sleep as a prerequisite) / / Symptom 4

[0083] (DSM-5 diagnostic criteria for depressive disorders, with prerequisites including psychomotor agitation or retardation) / / Symptom 5

[0084] (DSM-5 diagnostic criteria for depressive disorders, with a prerequisite including fatigue or lack of energy) / / Symptom 6

[0085] (DSM-5 diagnostic criteria for depressive disorders, with the prerequisite of feeling worthless or excessively guilty) / / Symptom 7

[0086] (DSM-5 diagnostic criteria for depressive disorders, the prerequisites of which include impaired thinking ability or difficulty concentrating) / / Symptom 8

[0087] (DSM-5 diagnostic criteria for depressive disorders, with a prerequisite including recurrent thoughts of death or suicide) / / Symptom 9

[0088] d. Define the overall logical relationship between these symptom premises:

[0089] (All symptoms under the DSM-5 diagnostic criteria for depressive disorders, operator, 'AND') / / This is a holistic logic: all symptoms in the list are in an "AND" relationship.

[0090] (DSM-5 diagnostic criteria for depressive disorders, the prerequisite for which is impaired social functioning) / / Other conditions that would otherwise cause functional impairment must be excluded.

[0091] The above demonstrates an example of a constructed diagnostic rule. It should be understood that there may be a one-to-one, one-to-many, or many-to-one relationship between diagnostic rules and diseases.

[0092] 3) Development of treatment guidelines

[0093] The development of treatment guidelines is generally more complex, as it defines a complete clinical pathway from diagnosis to treatment.

[0094] ① Guideline Structuring and Mapping: Complex natural language guidelines (such as the "Chinese Guidelines for the Prevention and Treatment of Depression") are decomposed into decision trees, and each node and branch is mapped to existing entities and relationships in the knowledge graph, or new entity relationships are created to represent the decision logic. ② Constructing Guideline Subgraphs: A treatment guideline can also be modeled as a complex subgraph and integrated into the knowledge graph.

[0095] a. Core: Create a treatment guideline entity (e.g., "Adult Major Depression Treatment Guideline V2022").

[0096] b. Define the applicable objects: Link them to the target disease entity (such as "major depressive disorder") through the guideline_applicable_disease relationship.

[0097] c. Enter recommended protocols: Create entities such as first-line treatment protocols and second-line treatment protocols, and link them to specific drugs (such as "sertraline" and "venlafaxine"), dosages, and treatment durations (defined through attributes). Then, connect these protocol entities to treatment guideline entities through the guideline_recommendation_protocol relationship.

[0098] d. Inclusion of contraindications and precautions: Link related diseases or patient conditions (e.g., “bipolar disorder”) through Protocol_Contraindications_Use relationship, and link drugs that have interactions (e.g., “monoamine oxidase inhibitors”) through Protocol_Precautions_Interaction relationship.

[0099] The above can be used to construct treatment guidelines. It should be understood that there is usually a many-to-many relationship between diseases and treatment recommendations (or treatment plans).

[0100] As described above, a fundamental transformation is achieved by embedding diagnostic rules and treatment guidelines into the knowledge graph as part of the construction process. Specifically, the knowledge graph is no longer merely a queryable database, but a knowledge engine embedded with clinical logic, capable of proactive logical reasoning. Centralized management and unified updates of all rules and guidelines ensure the consistency and authority of the outputs of all knowledge graph-based applications. When a user queries "What disease might a patient's symptoms indicate?", the system no longer simply lists related diseases, but activates these embedded rule subgraphs, matching the patient's symptoms with the rule premises, ultimately providing structured and confident diagnostic suggestions. Treatment recommendations follow the same principle. It should be noted that relying solely on knowledge graphs for reasoning to obtain auxiliary diagnostic results has limitations, such as fixed reasoning paths and a lack of flexibility.

[0101] The above introduces the construction of knowledge graphs based on mental illness data.

[0102] II. Label System Design

[0103] Based on the constructed knowledge graph, a tagging system can be designed. Specifically, the designed tagging system can include diagnostic tags and treatment suggestion tags. Diagnostic tags can be further subdivided into: 1) Disease diagnosis tags: diagnostic results for major disease types such as depressive disorders, anxiety disorders, and sleep disorders. 2) Severity tags: severity levels such as mild, moderate, and severe. 3) Comorbidity tags: comorbidity diagnoses of multiple coexisting diseases.

[0104] Treatment recommendation labels can include major treatment options such as medication, psychotherapy, and behavioral intervention.

[0105] It should be noted that the tags in the tagging system correspond to entities in the knowledge graph. For example, a mapping relationship can be established between each tag and a disease entity or treatment suggestion entity in the knowledge graph. Alternatively, the tag name can be directly determined as the name of the corresponding entity in the knowledge graph.

[0106] Furthermore, the labels in the labeling system are verifiable, which is reflected in: 1) Diagnostic basis: Each diagnostic label has a clear DSM-5 / ICD-11 diagnostic standard basis. 2) Symptom association: The association between the diagnostic label and specific symptoms. 3) Exclusion criteria: Clearly defined basis for excluding other diseases. 4) Verification path: Providing a specific path to verify the correctness of the diagnostic label. In short, verifiable labels have clear verification paths and medical basis, supporting the traceability of the reasoning process.

[0107] Based on the above, the design of the labeling system can be realized.

[0108] III. Synthesizing Diagnostic and Treatment Labeling Data Based on Knowledge Graphs and Tagging Systems

[0109] Figure 3 The document illustrates the process steps for automatically synthesizing diagnostic and treatment annotation data (or diagnostic and treatment samples) based on the constructed knowledge graph and designed tagging system. It should be understood that the entity executing this method can be any device, server, platform, or equipment cluster with computing and processing capabilities. Figure 3 The method shown in the figure includes the following process steps:

[0110] First, in step S310, a knowledge graph constructed based on mental illness data is used to identify several candidate symptom combinations (or candidate symptom combination sets). It should be understood that "several" in this text refers to one or more.

[0111] In one implementation, symptom combinations can be determined using real case data. It should be understood that real-world patient case data exists, and while this data has some reference value, its inconsistent quality makes it difficult to use directly as high-quality labeled data. Therefore, this proposal suggests using this data to construct high-quality diagnostic and treatment samples.

[0112] Specifically, for the first real case data, multiple standard symptom terms are identified to form the first candidate symptom combination. It should be noted that the first real case data refers to any one of the multiple case data under mental illness, and the "first" in "first real medical record data," and similar terms such as "second" and "third" elsewhere in the text, are used to distinguish similar items and do not have any ranking or other limiting function. Furthermore, the multiple standard symptom terms are entities in the knowledge graph.

[0113] There are several ways to link real case data to symptom-related entities in a knowledge graph.

[0114] In one embodiment, first real case data can be input into a trained entity recognition model to obtain multiple standard symptom terms as output. It should be understood that the training data of the entity recognition model may include labeled data pairs, such as (free text, standard term); the entity recognition model may be a lightweight model specifically designed for entity recognition, or it may be a model obtained by fine-tuning a pre-trained large model (such as BERT, BioBERT, etc.).

[0115] In another embodiment, symptom-related entities in the knowledge graph can be constructed as a dictionary, where each entry includes a standardized entity and its corresponding common expressions. Thus, the first set of real case data can be segmented, and the resulting segments can be matched with the dictionary to obtain the aforementioned standardized symptom terms.

[0116] From the above, we can obtain the candidate symptom combinations corresponding to the first real case data. Similarly, we can obtain candidate symptom combinations corresponding to one or more real case data sets. Furthermore, for the same real case data set, it can be repeatedly input into the entity recognition model to obtain candidate symptom combinations from each output, which are then categorized into the aforementioned candidate symptom combination set.

[0117] In another implementation, instead of relying on real case data, symptoms are extracted directly from the subgraphs containing the corresponding diagnostic rules in the knowledge graph to obtain candidate symptom combinations.

[0118] In one embodiment, symptoms under different modalities (such as language, vision, physiology, behavior, and scales) can be extracted from a subgraph corresponding to a diagnostic rule to form a corresponding candidate symptom combination.

[0119] In another embodiment, symptoms with temporal relationships can be extracted from the subgraph corresponding to a diagnostic rule to obtain candidate symptom combinations. It should be understood that temporal relationships can correspond to specific relationship types in a knowledge graph. Further, in one example, based on the relationship type "symptom A_follows_symptom B", symptom entities such as loss of interest and persistent low mood can be extracted from the subgraph corresponding to the major depressive disorder diagnostic rule. In another example, based on the relationship type "symptom A_occurs simultaneously with symptom B", symptom entities such as psychomotor agitation and flight of ideas can be extracted from the subgraph corresponding to the mania diagnostic rule.

[0120] It should be understood that a certain combination of symptoms can include both multimodal features and temporal features.

[0121] Based on the above, several candidate symptom combinations can be obtained using the knowledge graph.

[0122] Next, in step S320, for any first candidate symptom combination, the subgraph of the corresponding diagnostic rule in the knowledge graph is used to reason about the symptom combination to obtain a first diagnostic category label. It should be noted that the implementation method for this step can be flexibly selected as needed.

[0123] In one implementation, the first candidate symptom combination and disease diagnosis intent can be converted into a first graph query statement, and then this first graph query statement can be provided to the knowledge graph database, so that the graph database can use the subgraph data of the corresponding diagnosis rule in the knowledge graph to perform a query and return one or more disease entities as the first diagnosis class label.

[0124] Furthermore, in a specific embodiment, as mentioned above, the corresponding first diagnostic rule is explicit during the determination of the first candidate symptom combination. However, considering that the mapping relationship between diagnostic rules and diseases may be one-to-many, many-to-one, or one-to-one, it is proposed that the graph query statement may also include the identifier of the first diagnostic rule, thereby reducing the computational resources and time consumed by the query.

[0125] In another implementation, the diagnostic rules in the knowledge graph can be pre-converted into the form of "If-Then", and then a separate rule engine can be used to process the reasoning in this step.

[0126] In another implementation, graph-based computation or graph embedding representation-based semantic reasoning can be employed.

[0127] From the above, we can obtain the diagnostic labels corresponding to the first candidate symptom combination. For example, the first candidate symptom combination includes: persistent low mood, loss of interest, and difficulty falling asleep. The corresponding diagnostic labels may include: comorbid depression and insomnia, mild depression, and mild sleep disorder.

[0128] Then, in step S330, using the subgraph of the treatment guide corresponding to the treatment guide in the knowledge graph, reasoning is performed on the diagnostic tag and the first candidate symptom combination to obtain the first treatment suggestion tag. It should be noted that the implementation method for this step can be flexibly selected as needed.

[0129] In one implementation, the first candidate symptom combination and its diagnostic class label can be converted into a second graph query statement, and then this second graph query statement can be provided to the knowledge graph database, so that the graph database can use the subgraph data of the corresponding treatment guide in the knowledge graph to perform a query and return one or more treatment suggestion entities as the first treatment suggestion label.

[0130] In other implementations, the first treatment suggestion label can be obtained by constructing a rule engine for the treatment guidelines, or by performing graph-based algorithm computation or semantic reasoning based on graph embedding representation. For example, the first treatment suggestion label may include psychotherapy and drug therapy, etc.

[0131] Then, in step S340, the first candidate symptom combination, the first diagnostic category label, and the first treatment suggestion label are constructed into a first candidate sample to determine the diagnosis and treatment sample set.

[0132] In one embodiment, the first candidate sample can be directly used as a treatment sample and included in the treatment sample set.

[0133] In another embodiment, to further improve the quality of the final determined treatment sample set, it is proposed to conduct a quality assessment on the first candidate sample, and if the assessment is passed, it will be included in the treatment sample set; otherwise, it will be discarded.

[0134] In one specific embodiment, a first large model can be used to score the first candidate sample. If the score is greater than a predetermined threshold, the first candidate sample is included in the diagnostic sample set; otherwise, it is discarded. It should be understood that the first large model can be a high-performance commercial large model.

[0135] In one example, the scoring task instruction input into the first large model along with the first candidate sample could include: scoring the diagnosis and treatment plan, with a maximum score of 10. In another example, several scoring examples could also be included when inputting into the first large model along with the first candidate sample, for example, scoring examples with scores of 0, 2, 4, 6, 8, and 10, for the first large model to learn from.

[0136] In another specific embodiment, assuming that the same real case data is repeatedly input into the entity recognition model to obtain multiple different candidate symptom combinations, then the diagnosis and treatment labels (diagnosis labels and treatment suggestion labels) corresponding to these multiple different candidate symptom combinations can be voted on. Further, if the highest vote rate is greater than a predetermined threshold, the candidate sample containing the diagnosis and treatment label corresponding to the highest vote rate is included in the diagnosis and treatment sample set, and the others are discarded; otherwise, all are discarded.

[0137] The above allows us to determine the diagnostic and treatment sample set, thus enabling the automatic synthesis of diagnostic and treatment labeled data based on knowledge graphs and labeling systems. This diagnostic and treatment labeled data is obtained at a low cost and with high quality.

[0138] IV. Training-Assisted Diagnosis and Treatment Large Model

[0139] After automatically synthesizing high-quality diagnostic and treatment annotation data, it can be used to train a large-scale auxiliary diagnostic and treatment model. Figure 4 The flowchart illustrates the steps of the training method. The subject executing this method can be any device, platform, server, or equipment cluster with computing and processing capabilities. Figure 4 The training method shown includes the following steps:

[0140] Step S410: Transform the diagnosis and treatment sample set into a fine-tuning sample set containing inference chains using the second large model; Step S420: Perform supervised fine-tuning of the auxiliary diagnosis and treatment large model using the fine-tuning sample set; Step S430: Further fine-tune the auxiliary diagnosis and treatment large model using a reinforcement learning algorithm, with its reward score derived from the performance of the diagnosis and treatment large model on the diagnosis and treatment sample set.

[0141] The steps described above are explained in detail below:

[0142] First, in step S410, the diagnosis and treatment sample set is transformed into a fine-tuned sample set containing the inference chain using the second large model.

[0143] It should be noted that each diagnosis and treatment sample in the aforementioned sample set includes symptom combinations and corresponding diagnosis and treatment labels, but does not include the chain of thought (CoT), i.e., the reasoning steps. The chain of thought enables the interpretability of the diagnosis and treatment results. Therefore, this paper proposes to modify the sample set to include correct chain of thought, thereby providing a learning tool for the large-scale diagnostic and treatment model.

[0144] Specifically, for any first treatment sample in the treatment sample set, the second major model is used to process the first symptom combination in that sample to obtain the corresponding first inference chain and predicted treatment outcome. It should be understood that the second major model can be a commercially viable large model with good generalization performance, such as a commercial medical model. The second major model can be the same as or different from the first major model mentioned above.

[0145] Furthermore, if the predicted diagnosis and treatment result successfully matches the diagnosis and treatment label of the first diagnosis and treatment sample, the first symptom combination, the first inference chain, and the diagnosis and treatment label are constructed as the first fine-tuning sample and included in the fine-tuning sample set.

[0146] It should be noted that, regarding how to determine whether a match is successful, in one embodiment, named entity recognition can be performed on the predicted diagnosis result to obtain its hit label for the label system. Then, it can be determined whether this hit label covers the diagnosis label, or the coverage rate can be determined. If full coverage or a coverage rate greater than a predetermined threshold is determined, the match is considered successful; otherwise, the match fails. In another embodiment, the semantic similarity between the predicted diagnosis result and the diagnosis label corresponding to the first diagnosis sample can be calculated. If the semantic similarity is greater than a predetermined threshold, the match is considered successful; otherwise, the match fails.

[0147] Additionally, the second model can be used to infer the first treatment sample multiple times (e.g., 10 times), thereby determining whether each inference result is a successful match, and constructing corresponding fine-tuned samples if a match is successful. In this way, a richer set of fine-tuned samples can be obtained.

[0148] From the above, we can obtain a fine-tuned sample set containing the inference chain.

[0149] Next, in S420, the large-scale auxiliary diagnosis and treatment model is subjected to supervised fine-tuning using the fine-tuning sample set.

[0150] It should be noted that the training party has known the model parameters of the auxiliary diagnosis and treatment large model. For example, the auxiliary diagnosis and treatment large model can be a large model developed and pre-trained by the training party.

[0151] In one implementation, all parameters in the large-scale auxiliary diagnosis and treatment model can be fine-tuned.

[0152] In another implementation, in order to reduce computational costs and ensure good fine-tuning results, a parameter-efficient fine-tuning (PEFT) technique is proposed. For example, low-rank adaptation (LoRA) or adapter techniques can be used. For details, please refer to the relevant introductions in existing materials, which will not be elaborated here.

[0153] From the above, we can obtain a preliminary, finely tuned auxiliary diagnosis and treatment model, which has the basic ability to provide diagnosis and treatment results and reasoning chains.

[0154] Then, in step S430, the auxiliary diagnosis and treatment model is further fine-tuned using a reinforcement learning algorithm, and its reward score is derived from the performance of the diagnosis and treatment model on the diagnosis and treatment sample set.

[0155] It should be understood that reinforcement learning (RL) algorithms can employ proximal policy optimization (PPO) algorithms, actor-critic algorithms, and so on.

[0156] The following uses the PPO algorithm as an example to illustrate the execution of this step. Specifically,

[0157] 1) The auxiliary diagnostic model is used to process the symptom combinations sampled from the diagnostic sample set and output diagnostic actions. It should be understood that this diagnostic action includes predicting the diagnostic outcome and the inference chain.

[0158] 2) The reward score is determined by comparing the diagnostic and treatment actions with the corresponding diagnostic and treatment tags for the symptom combinations. It should be understood that the predicted diagnostic and treatment results and diagnostic and treatment tags in the diagnostic and treatment actions can be compared.

[0159] In one example, named entity recognition can be performed on the predicted diagnosis and treatment results. Furthermore, if the recognized entity covers the diagnosis and treatment label, the reward score can be set to 1; otherwise, it can be set to 0.

[0160] In another example, the semantic similarity between the predicted diagnosis and treatment result and the diagnosis and treatment label can be calculated. If the semantic similarity is greater than a predetermined threshold, the semantic similarity is used as the reward score; otherwise, -1 is determined as the reward score.

[0161] 3) Adjust the model parameters in the auxiliary diagnosis and treatment model with the goal of maximizing the expected value of the reward score. It should be understood that for the parameter tuning in this step, full parameter tuning or partial parameter tuning using the PEFT technique can be performed.

[0162] Based on the above, the RL algorithm can be used to further fine-tune the large-scale auxiliary diagnosis and treatment model, resulting in a final, more generalizable auxiliary medical model for practical use. It should be understood that the auxiliary medical model can be used by doctors (or other users) based on patient symptoms to invoke the model, thereby obtaining returned diagnostic results, treatment suggestions, and inference chains, which in turn assist in the final diagnosis and treatment.

[0163] In summary, the improved scheme disclosed in the embodiments of this specification first constructs a knowledge graph based on prior knowledge of mental illnesses (such as multi-source data, multimodal data, etc.), and then synthesizes labeled data for the diagnosis and treatment of mental illnesses based on this knowledge graph, thereby using this labeled data to train a large-scale auxiliary diagnosis and treatment model. Thus, 1) constructing a knowledge graph allows for the introduction of prior knowledge and the effective integration of complex relationships between mental illness data. 2) the knowledge graph can be updated according to updates to domain-related data, thereby resynthesizing some labeled data to retrain the large-scale auxiliary diagnosis and treatment model, enabling the diagnosis and treatment system to keep pace with knowledge updates. 3) the labeled data is automatically synthesized based on the knowledge graph, resulting in high quality and low cost. 4) the trained large-scale auxiliary diagnosis and treatment model provides highly accurate diagnosis and treatment results, and also produces inference chains along with the diagnosis and treatment results, making the diagnosis and treatment results verifiable and interpretable.

[0164] Corresponding to the above-mentioned methods for constructing labeled data and training large-scale models for assisted diagnosis and treatment, the embodiments of this specification also disclose a construction apparatus and a training apparatus. Specifically, as follows:

[0165] Figure 5 The device 500 for constructing a diagnostic sample includes the following functional modules:

[0166] The symptom combination determination module 510 is configured to determine several candidate symptom combinations using a knowledge graph constructed based on mental illness data. The diagnosis label determination module 520 is configured to, for any first candidate symptom combination, use a subgraph of the corresponding diagnostic rule in the knowledge graph to reason about the symptom combination and obtain a diagnostic label. The treatment label determination module 530 is configured to use a subgraph of the corresponding treatment guideline in the knowledge graph to reason about the first candidate symptom combination and the diagnostic label to obtain a treatment suggestion label. The treatment sample set determination module 540 is configured to construct a first candidate sample from the first candidate symptom combination, the diagnostic label, and the treatment suggestion label to determine the treatment sample set.

[0167] In one possible design, the mental illness data includes multimodal data, specifically including at least one of the following: language modality data, visual modality data, physiological modality data, behavioral modality data, and scale modality data.

[0168] In one possible design, the symptom combination determination module 510 is specifically configured to: for any first real case data, determine its corresponding multiple standard symptom terms to form the first candidate symptom combination; the multiple standard symptom terms belong to entities in the knowledge graph.

[0169] In one possible design, the symptom combination determination module 510 is specifically configured to: extract symptoms under different modalities from the subgraph corresponding to the first diagnostic rule in the knowledge graph to form the first candidate symptom combination.

[0170] In one possible design, the plurality of candidate symptom combinations includes a second candidate symptom combination, wherein the symptoms relate to one or more of the following modalities: language, visual, physiological, behavioral, and scale; and / or, the plurality of candidate symptom combinations includes a third candidate symptom combination, which includes historical symptoms and current symptoms.

[0171] In one possible design, the diagnostic labels include disease diagnosis labels, severity labels, and comorbidity labels.

[0172] In one possible design, the diagnostic tags and treatment suggestion tags correspond to entities in the knowledge graph.

[0173] In one possible design, the constructing device 500 further includes a scoring module configured to score the first candidate sample using a first large model, and to include the first candidate sample in the diagnostic sample set if the score is greater than a predetermined threshold.

[0174] Figure 6 The training device 600 for the large-scale auxiliary diagnosis and treatment model is shown, which includes the following functional modules:

[0175] The fine-tuning sample set determination module 610 is configured to transform the diagnosis and treatment sample set into a fine-tuning sample set containing an inference chain using the second large model. The diagnosis and treatment sample set adopts... Figure 5 The apparatus shown is used to determine the following: a supervised fine-tuning module 620 configured to supervise the fine-tuning of the auxiliary diagnosis and treatment model using the fine-tuning sample set; and a reinforcement learning module 630 configured to further fine-tune the auxiliary diagnosis and treatment model using a reinforcement learning algorithm, wherein the reward score is derived from the performance of the diagnosis and treatment model on the diagnosis and treatment sample set.

[0176] In one possible design, the fine-tuning sample set determination module 610 is specifically configured as follows: for any first diagnosis sample in the diagnosis sample set, the second large model is used to process the first symptom combination in the sample to obtain the corresponding first inference chain and predicted diagnosis result; if the predicted diagnosis result successfully matches the diagnosis label of the first diagnosis sample, the first symptom combination, the first inference chain and the diagnosis label are constructed as the first fine-tuning sample.

[0177] In one possible design, the reinforcement learning module 630 is specifically configured as follows: using the auxiliary diagnosis and treatment large model to process symptom combinations sampled from the diagnosis and treatment sample set, and outputting diagnosis and treatment actions; determining a reward score by comparing the diagnosis and treatment actions and the diagnosis and treatment labels corresponding to the symptom combinations; and adjusting the model parameters in the auxiliary diagnosis and treatment large model with the goal of maximizing the expected value of the reward score.

[0178] It should be noted that for a description of the above functional units, please refer to the relevant description of the process method in the foregoing embodiments.

[0179] In this specification, a large language model (or simply large model) is a natural language processing model based on deep learning technology. Its parameter count typically ranges from billions to hundreds of billions or even higher, possessing powerful language understanding and generation capabilities. Large language models can employ the Transformer architecture or its variants (such as GPT, BERT, etc.), which utilizes an attention mechanism to globally model sequential data, efficiently handling long-distance dependencies and thus performing exceptionally well in natural language tasks. Large language models learn the statistical features and semantic relationships of language through pre-training on large-scale corpora, giving them outstanding generalization capabilities. The core capabilities of large language models include, but are not limited to: understanding contextual semantics, generating coherent and grammatically correct text, performing logical reasoning, and handling multi-task scenarios. Their usage typically includes two modes: direct inference and fine-tuning. In direct inference mode, the user guides the large language model to generate specific outputs by designing prompts. Prompts can be task descriptions or instructions in text form, used to stimulate the large language model's semantic understanding and generation capabilities. In fine-tuning mode, large language models are further trained on small-scale datasets within a specific domain to optimize their performance on specific tasks. The powerful generalization capabilities and flexibility of large language models make them an important tool in the field of artificial intelligence, providing efficient and accurate solutions for automated text generation and understanding.

[0180] In some embodiments, large language models can also understand and generate data from other modalities (such as visual and audio data). In this case, large language models can also be called multimodal large language models (MLLMs). MLLMs provide a richer and more natural interactive experience by integrating multiple types of input and output, such as text, images, and sound. The core advantage of MLLMs lies in their ability to process and understand information from different modalities and fuse this information to complete complex tasks. For example, MLLMs can analyze an image and generate descriptive text, or generate a corresponding image based on a text description. This cross-modal understanding and generation capability makes MLLMs widely applicable across multiple fields.

[0181] It should be noted that the key technologies of large language models can be found in the detailed description in the paper "A Survey of Large Language Models" (paper number: arXiv:2303.18223v16, published on March 11, 2025, public link: https: / / doi.org / 10.48550 / arXiv.2303.18223), and will not be repeated here.

[0182] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform... Figure 3 or Figure 4 The method described.

[0183] According to another embodiment, a computing device is also provided, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement... Figure 3 or Figure 4 The method described.

[0184] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0185] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a diagnostic sample, comprising: Using a knowledge graph built from mental illness data, several candidate symptom combinations were identified; For any first candidate symptom combination, the symptom combination is inferred using the subgraph of the corresponding diagnostic rule in the knowledge graph to obtain a diagnostic label; Using the subgraph corresponding to the treatment guidelines in the knowledge graph, reasoning is performed on the first candidate symptom combination and diagnostic tags to obtain treatment suggestion tags; The first candidate symptom combination, diagnostic label, and treatment suggestion label are used to construct the first candidate sample to determine the diagnosis and treatment sample set.

2. The method according to claim 1, wherein, The mental illness data includes multimodal data, specifically including at least one of the following: language modality data, visual modality data, physiological modality data, behavioral modality data, and scale modality data.

3. The method according to claim 1, wherein, Using a knowledge graph built from mental illness data, several candidate symptom combinations were identified, including: For any first real case data, determine its corresponding multiple standard symptom terms to form the first candidate symptom combination; the multiple standard symptom terms belong to entities in the knowledge graph.

4. The method according to claim 1, wherein, Using a knowledge graph built from mental illness data, several candidate symptom combinations were identified, including: Symptoms under different modalities are extracted from the subgraphs of the knowledge graph corresponding to the first diagnostic rule to form the first candidate symptom combination.

5. The method according to claim 1, wherein, The plurality of candidate symptom combinations includes a second candidate symptom combination, wherein the symptoms involve one or more of the following modalities: language, visual, physiological, behavioral, and scale; and / or, The candidate symptom combinations include a third candidate symptom combination, which includes historical symptoms and current symptoms.

6. The method according to claim 1, wherein, The diagnostic labels include disease diagnosis labels, severity labels, and comorbidity labels.

7. The method according to claim 1, wherein, The diagnostic tags and treatment suggestion tags correspond to entities in the knowledge graph.

8. The method according to claim 1, wherein, After constructing the first candidate sample by combining the first candidate symptom combination, diagnostic label, and treatment suggestion label, the method further includes: The first candidate sample is scored using the first large model. If the score is greater than a predetermined threshold, the first candidate sample is included in the diagnostic sample set.

9. A training method for a large-scale auxiliary diagnosis and treatment model, comprising: The second major model is used to transform the diagnosis and treatment sample set into a fine-tuned sample set containing an inference chain, wherein the diagnosis and treatment sample set is determined by the method described in claim 1; The auxiliary diagnosis and treatment model is subjected to supervised fine-tuning using the fine-tuning sample set. The auxiliary diagnosis and treatment model is further fine-tuned using a reinforcement learning algorithm, and its reward score is derived from the performance of the diagnosis and treatment model on the diagnosis and treatment sample set.

10. The method according to claim 9, wherein, The second major model is used to transform the clinical sample set into a fine-tuned sample set containing inference chains, including: For any first medical sample in the medical sample set, the second large model is used to process the first symptom combination in the sample to obtain the corresponding first inference chain and the predicted medical result. If the predicted diagnosis and treatment result successfully matches the diagnosis and treatment label of the first diagnosis and treatment sample, the first symptom combination, the first inference chain, and the diagnosis and treatment label are constructed as the first fine-tuning sample.

11. The method according to claim 9, wherein, The large-scale auxiliary diagnosis and treatment model is further fine-tuned using reinforcement learning algorithms, including: The auxiliary diagnostic model is used to process symptom combinations sampled from the diagnostic sample set and output diagnostic actions. The reward score is determined by comparing the diagnostic and treatment actions with the diagnostic and treatment tags corresponding to the symptom combinations; The model parameters in the auxiliary diagnosis and treatment model are adjusted with the goal of maximizing the expected value of the reward score.

12. A device for constructing a diagnostic sample, comprising: The symptom combination determination module is configured to use a knowledge graph built based on mental illness data to determine several candidate symptom combinations; The diagnostic label determination module is configured to, for any first candidate symptom combination, use the subgraph of the corresponding diagnostic rule in the knowledge graph to reason about the symptom combination and obtain a diagnostic label. The treatment label determination module is configured to use the subgraph of the corresponding treatment guide in the knowledge graph to reason about the first candidate symptom combination and diagnostic labels to obtain treatment suggestion labels; The diagnostic sample set determination module is configured to construct a first candidate sample from the first candidate symptom combination, diagnostic label, and treatment suggestion label, for determining the diagnostic sample set.

13. A training device for a large-scale auxiliary diagnostic and treatment model, comprising: The fine-tuning sample set determination module is configured to transform the diagnosis and treatment sample set into a fine-tuning sample set containing an inference chain using the second large model, wherein the diagnosis and treatment sample set is determined using the apparatus of claim 12; The supervised fine-tuning module is configured to perform supervised fine-tuning of the auxiliary diagnosis and treatment model using the fine-tuning sample set. The reinforcement learning module is configured to use a reinforcement learning algorithm to further fine-tune the auxiliary diagnosis and treatment model, and its reward score is derived from the performance of the diagnosis and treatment model on the diagnosis and treatment sample set.

14. A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed in the computer, it causes the computer to perform the method according to any one of claims 1-11.

15. A computing device comprising a memory and a processor, wherein, The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-11.

Images