Semi-structured mental disorder interview method based on large model enhancement

Abstract

The invention discloses a semi-structured mental disorder interview method based on large model enhancement, and relates to the cross field of clinical psychology and artificial intelligence, and the method comprises the steps: extracting key information according to a medical text to construct an interview question set, generating a standardized question through a large model, and obtaining a standardized question set; the expression style is adjusted, and the problem quality is optimized in combination with a self-evaluation feedback mechanism; identifying similar points and identification points between symptoms by using a BERT classifier, and generating a decision tree structure in combination with an ICL technology and an XOT framework; a multi-Agent collaborative dialogue framework fused with an RAG mechanism is adopted, a large model and a simulated patient are used for multiple rounds of dialogues, context memory, fine-grained questioning and decision rule retrieval are processed respectively, and the inquiry process is guided; a mental disorder interview evaluation system is constructed, expert knowledge and large model ability are fused, interview content is subjected to multi-dimensional evaluation, and a diversified data set is generated through interview simulation. The method can effectively improve the efficiency and accuracy of mental disorder auxiliary diagnosis.

Description

Technical Field

[0001] The present invention relates to the cross - field of clinical psychology and artificial intelligence, and particularly to a semi - structured interview method for mental disorders enhanced by large models. Background Art

[0002] Mental disorders are common and serious public health problems that affect an individual's emotions, thinking, behavior, and social functions, and are characterized by a long course, high recurrence rate, and significant individual differences. In clinical practice, the diagnosis of mental disorders mainly relies on professional physicians through face - to - face interviews with patients, comprehensively judging their chief complaints, behavioral characteristics, medical history information, and environmental factors. Currently, commonly used auxiliary tools include SCID (Structured Clinical Interview) and various scales designed for specific diseases such as depression, anxiety, bipolar disorder, schizophrenia, etc., such as HAMD, PHQ - 9, GAD - 7, etc. These tools have certain value in improving diagnostic consistency, but there are problems such as high artificial execution cost, rigid evaluation process, and difficulty in expanding to combined multi - disease assessment.

[0003] In the field of mental health, large models are currently mainly applied to tasks such as psychological dialogue generation, psychological counseling assistance, emotion recognition, and disease classification. In contrast, the automated interview technology for assisting in the diagnosis of mental disorders is still in its infancy, and its tasks include question generation, information collection, differential diagnosis, etc. However, the current technology still faces challenges in realizing the above functions:

[0004] 1. There is a lack of an effective mechanism to introduce structured clinical medical knowledge into the large - model reasoning process, resulting in the model being difficult to accurately distinguish the subtle differences between similar symptoms and insufficient diagnostic interpretability.

[0005] 2. Existing dialogue generation methods lack the ability to adapt to the questioning style and patient status, and are prone to problems such as inappropriate wording or logical jumps, affecting the coherence of the interview and the patient's cooperation.

[0006] Therefore, it is necessary to construct a semi - structured interview method for mental disorders that integrates large - language models and structured medical knowledge, has multi - module collaborative capabilities, can generate standardized questions, and realizes interpretable diagnostic reasoning, so as to improve the efficiency of medical interviews and the accuracy of diagnosis, and meet the clinical needs of intelligent assisted mental illness assessment. Summary of the Invention

[0007] The present invention aims to provide a semi-structured interview method for mental disorders enhanced by large models, which solves the problems existing in the prior art, such as lack of pertinence in questions, lack of interpretability in diagnostic reasoning, single dialogue structure, and imperfect interview evaluation system. By introducing the latest clinical guidelines and structured medical knowledge, combining with large language models to generate standardized questions, and using a multi-Agent collaborative framework to achieve dynamic follow-up and decision-making reasoning, while constructing an evaluation mechanism for interview content, the accuracy, coherence, and clinical adaptability of the automatic interview in the auxiliary diagnosis of mental disorders are improved.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A semi-structured interview method for mental disorders enhanced by large models, including:

[0010] 1. Interview question generation, specifically including:

[0011] Using a small model to extract diagnostic criteria, related symptoms, and differential diagnosis information from clinical guidelines, standard codes, and electronic medical records to obtain the core medical features required for interview questions;

[0012] Based on the extracted medical features, using a large model to generate a set of standardized interview questions that meet the requirements of clinical guidelines;

[0013] For the set of interview questions, identifying possible sensitive expressions through style-aware transformation, and standardizing the terms using the DSM-5 statistical manual;

[0014] Constructing a self-evaluation mechanism to analyze the expression quality, safety, and adaptability of the interview questions, adjusting the questions according to the feedback, and optimizing them cyclically.

[0015] 2. Diagnostic decision tree generation, specifically including:

[0016] Obtaining the case symptom description text of the medical record and clinical guidelines, using a BERT classifier to identify the similarities and differences of symptoms to obtain symptom labels;

[0017] Through ICL technology, inputting the symptom labels and case context to obtain the reasoning sequence structure of the case;

[0018] According to the symptom attribute labels and reasoning sequence structure, constructing an interpretable diagnostic decision tree with the help of the XOT framework, and recombining the interview questions;

[0019] For multiple case contents, generating multiple possible reasoning results for each, and selecting those with higher uncertainty for manual verification.

[0020] 3. Multi-Agent collaborative dialogue framework, specifically including:

[0021] Obtain the initial information of the patient, retrieve the relevant diagnostic decision tree, and use the large model to conduct multiple rounds of conversations with the patient based on the diagnostic decision tree process;

[0022] Through the context understanding and memory Agent, store and retrieve the patient background information accumulated during the interview;

[0023] Through the fine-grained questioning Agent, analyze the current diagnostic path, and generate refined questions for the nodes that are difficult to distinguish based on the context information;

[0024] Through the rule retrieval Agent, dynamically retrieve the standard clinical guidelines and differential diagnosis rules, and provide authoritative basis at the decision-making points;

[0025] Construct a feedback mechanism to evaluate the logic, rationality, and coherence of the conversation content, and optimize the subsequent questioning strategy.

[0026] 4. Mental disorder interview evaluation system, specifically including:

[0027] Simulate the roles of the interviewer and the interviewee, and evaluate from two aspects: the interview communication effect and the clinical decision-making effect;

[0028] Combine expert knowledge and large model analysis to construct a hybrid evaluation method to comprehensively evaluate the interview quality from multiple dimensions;

[0029] Measure the accuracy of symptom recognition and communication effect in the interview through traditional evaluation indicators such as precision, recall, and F1 score;

[0030] Measure the effectiveness of the interview and the emotional needs of the patient through artificial indicators such as accuracy, security, consistency, bias recognition, transparency, interpretability, and empathy level;

[0031] Based on the mental disorder interview evaluation system, generate a conversation dataset for mental disorder interviews.

[0032] The semi-structured mental disorder interview method enhanced by the large model proposed by the present invention has the following beneficial effects:

[0033] Improve the efficiency and standardization level of mental disorder interviews: This method can automatically generate questions that meet clinical specifications, replace the traditional way of manually writing questionnaires, greatly reduce the time cost required for medical staff to prepare and conduct interviews, make the mental disorder screening and diagnosis process more efficient and standard unified, and is especially suitable for primary medical institutions and large-scale screening scenarios.

[0034] Enhance patients' cooperation and interview experience: The expression style of interview questions can be adaptively adjusted according to the patient's background, avoiding rigid and stimulating language, reducing patients' defensive emotions and resistance to interviews, enhancing patients' subjective cooperation level and willingness to disclose information, and facilitating the acquisition of more comprehensive and accurate disease information.

[0035] Improve the scientificity and accuracy of clinical diagnosis: By systematically identifying the similarities and differences between symptoms and generating an interpretable diagnostic path, it assists doctors in accurately differentiating among multiple possibilities, thereby reducing misdiagnosis and missed diagnosis, especially applicable to patient groups with complex symptom manifestations or a high prevalence of comorbidities.

[0036] Ensure the safety and ethical compliance of automatically generated interview content: Introduce a content evaluation mechanism to continuously review and optimize the automatically generated questions, prevent inappropriate word usage, privacy-involved or unethical questioning methods, and ensure that the output complies with medical ethics and industry regulatory requirements, providing a reliable guarantee for the implementation of large models in medical scenarios.

[0037] Facilitate doctor-patient communication and decision-making sharing: The generated questions have a clear structure and reasonable logic, making it easy for patients to understand and provide accurate answers. At the same time, providing interpretable diagnostic bases also enables doctors and patients to jointly participate in the diagnosis and treatment decision-making process, enhancing the trust relationship and interaction quality between doctors and patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] To more clearly illustrate the technical solutions of the present invention, the following will briefly introduce the drawings required in the embodiments:

[0039] Figure 1 It is a flowchart of the semi-structured mental disorder interview method enhanced by a large model according to the present invention;

[0040] Figure 2 It is a flowchart of the automatic generation of interview questions according to the present invention;

[0041] Figure 3 It is a flowchart of the diagnosis reasoning integrating knowledge according to the present invention;

[0042] Figure 4 It is a flowchart of the multi-Agent collaborative dialogue framework according to the present invention;

[0043] Figure 5 It is a flowchart of the mental disorder interview evaluation system according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. The purpose of the present invention is to provide a semi-structured interview method for mental disorders enhanced by large models, which can effectively improve the interview efficiency, information acquisition quality, and diagnostic reasoning ability in the process of mental disorder assisted diagnosis. This method comprehensively utilizes clinical medical knowledge and the capabilities of large language models to achieve dynamic questioning, interpretable diagnostic reasoning, and secure assessment of interview content, and is applicable to various application scenarios such as hospital outpatient clinics, primary health service centers, and remote psychological intervention platforms.

[0045] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0046] As Figure 1 shown, the semi-structured interview method for mental disorders enhanced by large models provided by the present invention includes:

[0047] Step S1: Interview question generation. According to the latest clinical guidelines, electronic medical records, and standard coding, extract key information and features to construct an interview question set; generate standardized questions through a large model and adjust the expression style in combination with the patient's background; adopt a style-aware transformation technique to avoid sensitive expressions, and optimize the question quality in combination with a self-assessment feedback mechanism.

[0048] As Figure 2 shown, step S1 specifically includes:

[0049] Step S11: Collect text data from sources such as electronic medical records, clinical case reports, and medical guideline documents. To ensure the medical validity of diagnostic questions, this step uses a structured small model (lightweight BERT model) to perform fine analysis and extraction operations on the text data, including: collecting multi-source text data from electronic medical records, typical clinical case reports, and DSM / ICD clinical guideline documents; encoding the above text using the lightweight BERT model; automatically identifying and extracting the core diagnostic criteria (such as DSM-5 coding), symptom characteristics (subjective feelings, objective behaviors), disease exclusivity or co-occurrence indicators involved in the text; aligning and mapping the extraction results with the DSM / ICD standard coding table to achieve semantic standardization and consistency, facilitating tracking and verification; organizing the output results in a structured JSON form to provide an interpretable semantic basis for subsequent question generation.

[0050] Step S12: After obtaining medical features, call a large language model (LLM) and design in combination with Chain-of-Thought prompting to guide the model to generate interview questions that meet clinical specifications one by one in a way with a logical progressive structure. Each question needs to cover one or more specific diagnostic dimensions to ensure comprehensive coverage of key symptoms and discriminant factors; the generated questions follow the expression habits in medical guidelines, such as using clinical standard expressions like "Have you ever had...", "Have you often felt... in the recent period", etc.; the content of the questions needs to avoid leading, emotionally manipulative, or suggestive wording to ensure neutrality and clinical ethics; all generated questions are stored in a structured JSON format for subsequent sorting, style adjustment, and follow-up question generation. The output of this module is an initial set of interview questions, including several standardized questions (Q1, Q2,..., Qn), which serve as the input for the subsequent optimization process.

[0051] Step S13: For the preliminary interview questions generated by the large model, this step aims to make personalized adjustments to the question expression style and ensure that the content meets medical specifications and safety requirements. This step includes the following processes:

[0052] Automatically adjust the expression of the questions according to different patient backgrounds (including age levels, education levels, cultural environments, etc.), and select a language style that is easier to understand and accept;

[0053] Introduce the terms and expressions in the DSM-5 statistical manual to standardize the questions and ensure compliance with the internationally common diagnostic criteria for mental disorders;

[0054] Based on the DSM-5 differential diagnosis manual, analyze whether the current questions align with the required differential diagnosis points, such as whether they cover important contents such as exclusive symptoms and co-occurring symptoms;

[0055] Use a taboo vocabulary list to screen the question text, identify and replace words that do not conform to ethics or may cause discomfort, and ensure the safety and compliance of the interview process;

[0056] The adjusted questions retain the original medical semantic basis and are more adaptable and humanistic in expression, facilitating patient understanding and cooperation.

[0057] Step S14: After question generation and style adjustment, this step conducts a preliminary evaluation of the expression quality, safety, and style suitability of the questions through experts or models. The evaluation process includes whether the language is appropriate, whether there are ambiguous or risky words, and whether the corresponding background adaptation requirements are met. Through multiple rounds of iteration, the consistency, rationality, and clinical applicability of question generation are gradually improved. This evaluation and optimization process can be completed through the cooperation of manual and automated means to achieve a closed-loop optimization of interview question generation.

[0058] Step S2: Generate a diagnostic decision tree. Obtain the symptom description texts of medical records and clinical guidelines, use a BERT classifier to identify the similarities and differentiating points between symptoms, and generate a diagnostic decision tree by combining the ICL technique and the XOT framework. Learn differential diagnosis knowledge based on clinical guidelines, generate a diagnostic reasoning path, recombine the interview questions, and achieve interpretable diagnosis and treatment decisions to improve the diagnostic accuracy.

[0059] As Figure 3 shown, step S2 specifically includes:

[0060] Step S21: First, use a BERT-based symptom recognition module to process the symptom description texts from the Unlabeled Dataset Pool. This module is trained through contrastive learning and can identify features with "differentiating capabilities" in the symptom dimension (i.e., key symptoms that are valuable for differentiating different diseases, called differentiating points), as well as non-specific co-occurring symptoms lacking differentiating capabilities (called similarity points). The result Pred SLM output is a structured label set, indicating the attribute category of each symptom entry, serving as the input basis for the subsequent reasoning structure.

[0061] Step S22: Introduce the ICL (In-Context Learning) mechanism to understand the reasoning structure of clinical experts during the diagnosis process. ICL is essentially a few-shot learning method driven by a large language model, and its effectiveness depends on the provided context prompt combinations. In this embodiment, the prompt content of the ICL module includes: (1) labeled typical cases, including symptom lists and final diagnosis results; (2) diagnostic procedures for specific diseases in clinical guidelines; (3) disease feature paths described in medical literature. Input the symptom labels identified by Pred SLM together with the above context into the ICL model, which generates the reasoning sequence structure of this case in reasoning, such as determining whether the mood is elevated first, and then determining whether the duration meets the standard, etc. Finally, output Pred ICL as a multi-level reasoning sequence structure to guide the path distribution of the subsequent tree model.

[0062] Step S23: According to the aforementioned symptom attribute tags and the inference order structure, use the XOT (Exploratory Option Tree) framework to construct an interpretable diagnostic decision tree, and recombine the problems in step 101 according to the decision tree. XOT is a structured path simulation mechanism that can map symptom attributes to different branch judgment nodes in a tree structure and form a path leading to a specific diagnostic label according to different symptom combinations. Each non-leaf node of this decision tree represents the judgment condition of a "discrimination point", and the path order is constructed based on the recommended logical sorting of Pred ICL For example, taking an example structure, the first-level node of the tree is "Whether there is mania", if not, then enter "Whether there is long-term depression", if yes, then further judge "Whether there is an increase in activities", and finally lead to diagnostic labels such as "depressive disorder" or "bipolar disorder". Pred XOT is the instantiated tree model of this structured decision path and can be used to simulate the automatic inference of samples.

[0063] Step S24: Based on the constructed decision tree structure Pred XOT , perform simulated inference on unlabeled samples to evaluate the judgment stability of the model on this sample, and introduce an uncertainty index for sample screening. Uncertainty reflects the degree of diagnostic divergence of the model when facing the current symptom combination and is the key basis for measuring the "diagnostic difficulty" and "information value" of this sample. Specifically, for each sample to be evaluated, generate multiple possible inference results through different paths, multiple models or different context settings (such as multiple groups of ICL prompts) in the XOT structure. Each inference will produce a diagnostic result. Suppose a total of m prediction results (answers) are obtained, and there are h unique diagnostic conclusions (unique_answer) among them, then the calculation formula for uncertainty (abbreviated as u) is: u = h / m.

[0064] Among them, m represents the total number of diagnostic answers generated by the model on the current sample (including duplicates), and h represents the number of distinct ones among these answers. This formula represents the proportion of unique answers in all predictions and is used to characterize the degree of inference divergence: when u → 1, it means that the model has a high degree of diagnostic divergence for this sample, and the predictions are inconsistent, reflecting a high degree of uncertainty in the diagnosis of this sample; when u → 0, it means that multiple models / paths tend to be consistent and the judgment is stable.

[0065] Step S25: According to the uncertainty scoring results in step 1024, all unlabeled samples are sorted from high to low, and the top K samples are selected to enter the manual review process. The principle of selecting Top-K is that under the premise of limited resources, samples with the highest diagnostic ambiguity and representativeness are preferentially screened, so as to maximize the efficiency of model update. Sorting from high to low represents the strength of the "challenging" nature of the samples for the current model. The higher the ranking, the more it indicates that the internal diagnostic logic of the model is unstable or the current knowledge boundary is not covered, thus requiring manual supplementation of diagnostic path information. Manual review will be combined with the guidelines for differential diagnosis of mental disorders, and professionals will authoritatively confirm the symptom classification, decision-making path, and final diagnosis results, and incorporate them into the Labeled Dataset as new samples for supervised learning.

[0066] Step S26: Jointly update the core modules, absorb the feedback information from the previous round of manual annotation and use it for the next round of inference optimization. First, the SLM model will introduce new labeled samples for retraining to enhance its ability to recognize boundary symptom features, especially to improve the judgment sensitivity to "weak discriminative points". Second, the context prompt content of the ICL module will be expanded, adding the path information of successful inferences in this round of annotation cases to further enrich the diversity and clinical consistency of context reasoning. Finally, the XOT decision tree structure will perform operations such as pruning, reordering, or adding new branches to the original tree model according to the real inference paths in the new samples to ensure that the path structure is more stable, reasonable, and generalizable.

[0067] Step S27: Redeploy the updated model, return to step 1021, perform symptom recognition and inference judgment on new unlabeled samples, enter the next round of diagnostic inference process integrating knowledge, construct a closed-loop optimization structure, and gradually improve the overall diagnostic ability.

[0068] Step S3: A multi-Agent collaborative dialogue framework conducts multiple rounds of conversations with the patient based on patient information, using a diagnostic decision tree and a large model; adopts a multi-Agent framework with RAG fusion to separately handle context memory, fine-grained questioning, and decision rule retrieval; introduces a real-time evaluation mechanism to adjust the dialogue strategy and improve the coherence and diagnostic reliability of the interview.

[0069] As Figure 4 shown, step S3 specifically includes:

[0070] Step S31: First, preprocess based on the patient input information (such as chief complaint, medical history, symptoms, etc.), and retrieve the diagnostic decision tree that matches this type of patient from the knowledge base. Subsequently, conduct a structured multi-round conversation with the patient according to the retrieved decision tree path. Diagnostic path judgment task: Based on the patient's current information and combined with the reasoning ability of the large model, judge the next node that should be taken in the current node in the decision tree structure, that is, judge which type of symptom path or diagnostic branch the patient conforms to, so as to achieve dynamic path selection. Node content-driven question generation task: After clarifying the current node, extract the clinical question template or key symptom questions associated with this node, hand them over to the large model for semantic optimization and refinement, generate natural language questions in combination with the patient's context, and then interact with the patient to obtain the next piece of information.

[0071] Step S32: Introduce the "Context Understanding and Memory Agent" to summarize the patient's disease information in real time during the interview process and construct a structured context memory cache. This Agent calls the large language model and cooperates with a carefully designed prompt template for the disease summary task to ensure the semantic deep understanding and induction ability of multi-round natural language conversations. Through the prompt, guide the model to perform a summary processing of the historical conversation, store the extracted main symptom clues, the background medical history described by the patient, subjective feelings, etc. in a structured list form, and provide the context jointly constructed by the current "short-term memory" (such as the content of the recent 3-5 rounds of conversations) and the "long-term background" (such as the initial chief complaint and medical history) in each interrogation reasoning call. In this way, the model can call the complete disease chain information when making inferences and asking follow-up questions, so as to make reasonable judgments at the decision-making point, avoid information breaks and repeated questions, and significantly improve the coherence and professionalism of the conversation.

[0072] Step S33: Mobilize the "Fine-grained Follow-up Agent" to analyze the branch points in the current diagnostic path and judge whether there is a risk of "insufficient information" based on the known information. If, during the process of advancing along the path of the XOT decision tree, a certain key node lacks necessary discriminative symptom information (such as the lack of features such as "whether the mood swing is related to the circadian rhythm"), this Agent will call the aforementioned context memory and combine the current path position to automatically generate clinically meaningful fine-grained follow-up questions. Such follow-up questions include not only "yes / no" judgment questions, but also open-ended questions, aiming to obtain more accurate variables such as symptom expression, frequency, duration, or degree of impact. This Agent ensures that the judgment basis for each decision point has traceable clinical evidence support, so as to construct a decision-making chain that meets the guideline standards and avoid reasoning path breaks or misdiagnosis.

[0073] Step S34: Invoke the "Rule Retrieval Agent" to provide professional knowledge support for the current decision point. This Agent operates based on the RAG framework. The literature retrieval part is limited within the scope of standard clinical guidelines and differential diagnosis databases to ensure the authority and practicality of the returned content. In terms of technical implementation, this module first constructs a query vector through the semantics of the current node, and then calculates the semantic relevance with each paragraph in the knowledge base using cosine similarity, and selects the Top-K paragraphs as candidate reference texts. Subsequently, the language generation module integrates this content to output highly condensed and targeted knowledge prompts, which are used to assist in constructing fine-grained follow-up questions or directly presented to the doctor side for reference to support their understanding of the standard basis behind the judgment. This module significantly enhances the medical orientation of the interrogation process and the clinical interpretability of the reasoning process.

[0074] Step S35: Set up a feedback mechanism to evaluate the structural rationality of the current conversation in real time. This module conducts a coarse-grained evaluation of the language consistency, semantic fluency, and information integrity of the multi-round conversation. However, by providing feedback results, it can indirectly optimize the subsequent follow-up questioning strategy, thereby improving the overall conversation quality and patient adaptability, and preventing problems such as interrogation fatigue or structural chaos.

[0075] Step S4: Construct a mental disorder interview assessment system. Combine expert knowledge and large model analysis to evaluate the accuracy, safety, and ethics of the interview content; generate a dataset through interview simulation to provide reliable data support for academic research, clinical training, and model development.

[0076] As Figure 5 shown, Step S4 specifically includes:

[0077] Step S41: Enter the simulation interview stage. The simulation interviewer (Patient Actor) constructs a specific mental disorder scenario in combination with the disease scenario script (Scenario Pack). The interviewees are randomly assigned intern doctors or large language models (LLMs). The interview method is real-time text conversation. All interactions during the simulation interview are recorded and archived in structured text form (Consultation Transcript). It should be emphasized that the identity of the interviewee is set in a blind state, that is, the simulation interviewer does not know whether the person on the other side is a human or a model, to ensure the naturalness of the interview process and the objectivity of the results.

[0078] Step S42: Enter the stage of analyzing interview results, focusing on evaluating the communication skills and clinical decision-making abilities of the simulated interviewers. For the analysis of communication effectiveness, clinical interview assessment systems such as GMCPQ, PACES, and PCCBP are used to judge whether the interview achieves empathy, whether the structure is clear, etc.; while for the assessment of clinical decision-making ability, the simulated patient (Patient Actor) perceives the quality of the questions, judgment paths, and communication rationality of the interviewee in real time during the interview, and evaluates its effectiveness in recognition diagnosis, symptom understanding, and personalized expression in combination with the script.

[0079] Step S43: Introduce senior clinical experts and psychiatrists to participate in the manual assessment of interview quality. The experts analyze based on three types of core materials, including: interview script and standard answers (Scenario Pack + Ground Truth), actual interview text record (Consultation Transcript), and OSCE structured feedback questionnaire (Post-Questionnaire). The evaluation indicators cover multiple dimensions, namely accuracy, safety, consistency, bias detection, transparency, interpretability, and empathy level, to ensure that the model or intern not only provides professional judgments during the interview, but also has ethical care and patient adaptability.

[0080] Step S44: Automatically complete the structured quantitative analysis and introduce traditional evaluation indicators to accurately evaluate the interview effect. First, compare the diagnosis results identified in the interview with the standard answers, and calculate precision, recall, and F1 score to quantify the accuracy of disease recognition; secondly, match the decision tree path taken by the model (or intern) during the interview with the "standard diagnosis path" generated by the guidelines, evaluate the rationality and integrity in the clinical reasoning logic, and also perform accuracy analysis at the path level through the above three indicators (Precision / Recall / F1).

[0081] Step S45: Based on the above interview assessment system, comprehensively score all interview records, select the dialogue data with high scoring performance, coherent content, and strong clinical significance, and construct a high-quality interview dialogue dataset for mental disorders. This dataset retains key information such as dialogue context, patient background information, and clinical judgment basis, and is tagged to provide standardized and multi-level text resource support for clinical training, AI-assisted interrogation system optimization, and mental psychology research.

[0082] In the foregoing, specific embodiments of the present invention have been described with reference to the accompanying drawings. However, those of ordinary skill in the art can understand that various changes and substitutions can be made to the specific embodiments of the present invention without departing from the spirit and scope of the present invention. These changes and substitutions all fall within the scope defined by the claims of the present invention.

Claims

1. A semi-structured mental disorder interview method enhanced by large models, characterized in that, It includes the following steps: S1: Interview question generation: Extract key information and features based on the latest clinical guidelines, electronic medical records, and standard coding, and construct an interview question set; S2: Diagnostic decision tree generation: Obtain the symptom description texts of medical records and clinical guidelines, use a BERT classifier to identify similarities and differentiating points between symptoms, combine the ICL technology and the XOT framework to generate a diagnostic decision tree, and reorganize the interview questions according to the decision tree structure; S3: Multi-Agent dialogue framework: Based on patient information, use the diagnostic decision tree and a large model to conduct multiple rounds of conversations with the patient in simulation. Adopt a multi-Agent framework integrated with RAG to separately process context memory, fine-grained questioning, and decision rule retrieval to obtain an interview record and a diagnosis result of mental disorders; S4: Mental disorder interview evaluation system: Construct a mental disorder interview evaluation system, combine expert knowledge and large model analysis to evaluate the accuracy, safety, and ethics of the interview record, etc., and generate a data set through interview simulation.

2. The semi-structured mental disorder interview method enhanced based on a large model according to claim 1, wherein In the step S1, the interview question generation specifically includes: Use a small model to extract diagnostic criteria, related symptoms, and differential diagnosis information from clinical guidelines, standard coding, and electronic medical records to obtain the core medical features required for interview questions; Based on the extracted medical features, use a large model to generate a standardized interview question set that meets the requirements of clinical guidelines; For the interview question set, identify possible sensitive expressions through style-aware transformation, and standardize the terms using the DSM-5 statistical manual; Construct a self-evaluation mechanism to analyze the expression quality, safety, and adaptability of the interview questions, adjust the questions according to the feedback, and cycle for optimization.

3. The semi-structured mental disorder interview method enhanced based on a large model according to claim 1, characterized in that, In the step S2, the diagnostic decision tree generation specifically includes: Obtain the case symptom description texts of medical records and clinical guidelines, use a BERT classifier to identify similarities and differentiating points of symptoms to obtain symptom labels; Through the ICL technology, input the symptom labels and case context to obtain the inference sequence structure of the case; According to the symptom attribute labels and inference sequence structure, construct an interpretable diagnostic decision tree with the help of the XOT framework, and reorganize the interview questions.

4. The semi-structured mental disorder interview method enhanced based on a large model according to claim 1, characterized in that, In the step S3, the multi-Agent collaborative dialogue framework specifically includes: Obtain the initial patient information, retrieve the relevant diagnostic decision tree, and use a large model to conduct multiple rounds of conversations with the patient based on the diagnostic decision tree process; Store and retrieve the patient background information accumulated during the interview through the context understanding and memory Agent; Analyze the current diagnostic path through the fine-grained questioning Agent, and generate refined questions for the nodes that are difficult to distinguish based on the context information; Dynamically retrieve the standard clinical guidelines and differential diagnosis rules through the rule retrieval Agent to provide an authoritative basis at the decision point; Construct a feedback mechanism to evaluate the logic, rationality, and coherence of the conversation content, and optimize the subsequent questioning strategy.

5. The semi-structured mental disorder interview method enhanced based on a large model according to claim 1, characterized in that, In the step S4, the mental disorder interview evaluation system specifically includes: Simulate the roles of interviewer and interviewee, and conduct evaluations from two aspects: the interview communication effect and the clinical decision-making effect. Combining expert knowledge and large model analysis, construct a hybrid evaluation method to comprehensively evaluate the quality of interviews from multiple dimensions; Measure the accuracy of symptom recognition and communication effectiveness in interviews through traditional evaluation metrics such as precision, recall, and F1 score; Measure the effectiveness of interviews and the emotional needs of patients through artificial metrics such as accuracy, security, consistency, bias recognition, transparency, interpretability, and empathy level; Generate a dialogue dataset for mental disorder interviews based on the mental disorder interview evaluation system.

Images