Depression auxiliary evaluation method based on large language model agent

Abstract

The invention discloses a depression auxiliary evaluation method based on a large language model agent. The depression auxiliary evaluation method is characterized by comprising the following steps: a) performing structured processing on a Hamilton depression scale; b) constructing a dynamic questioning agent based on a large language model, and realizing dynamic selection and adjustment of theme questions; c) constructing a multi-dimensional scoring agent to obtain sub-item scores of the depressive symptoms; and d) constructing a clinical decision-making agent, generating risk early warning, giving personalized diagnosis and treatment and intervention suggestions and the like. Compared with the prior art, the method has the advantages that real-time interaction in a natural language form is realized, a real interview process is simulated, the method is closer to psychological consultation practice, user experience and evaluation reliability and validity are enhanced, user language behaviors and scoring results can be analyzed in real time, high-risk signals such as self-injury tendency and severe depression can be identified, and the method is suitable for popularization and application. Personalized diagnosis and treatment suggestions and intervention measures are output, intelligent assistance is provided for clinical decision making, and the intelligent level of mental health screening and management is improved.

Description

Technical Field

[0001] The present invention relates to the field of intersectional technology between artificial intelligence and medical health, and specifically to a method for auxiliary assessment of depression based on a large language model intelligent agent. Background Art

[0002] Depression, a highly prevalent mental illness, has become a major burden globally, severely impacting public mental health and social productivity. Effective depression assessment and intervention rely on clinicians' accurate assessment of patients' mental states. Currently, the Hamilton Depression Rating Scale (HAM-D) is widely used in psychiatric clinical practice as a standard tool for quantitatively assessing depression severity. This scale uses several items to score multiple dimensions, including mood, behavior, sleep, and weight, to assist clinicians in assessing depression severity and treatment effectiveness. However, traditional depression assessment methods often rely on manual interviews or static questionnaires, in which physicians or investigators ask patients item-by-item questions and record their subjective responses. This approach has several limitations: First, the assessment process is highly manual, time-consuming, susceptible to subjective judgment, and lacks consistency. Second, static questionnaires lack the ability to understand user semantic expressions and context, making it difficult to capture subtle changes in mental state. Third, traditional methods often fail to accurately translate to standard scoring systems when dealing with complex language expressions or ambiguous emotional descriptions, compromising the objectivity and accuracy of the final assessment results.

[0003] With the rapid development of artificial intelligence (AI), particularly the rise of large language models (LLMs) in recent years, natural language processing (NLP) capabilities have achieved significant breakthroughs. LLMs possess deep semantic understanding, contextual awareness, and language generation capabilities, enabling them to simulate human conversational behavior and autonomously handle complex language tasks. Intelligent agents built on these models have demonstrated broad application prospects in fields such as intelligent customer service, medical Q&A, and psychological support. In the field of mental health assessment, the use of LLMs for automated, intelligent question-answering and scoring has become a current research hotspot.

[0004] While some attempts have been made to incorporate natural language processing into psychological assessment systems, such as through sentiment analysis and keyword extraction to aid in determining a user's emotional state, these systems still face challenges such as crude methods, limited scoring dimensions, and a lack of professional medical knowledge. Furthermore, these systems are often unable to achieve dynamic conversational interaction with users and lack the ability to deeply model complex semantic structures and psychological characteristics. Summary of the Invention

[0005] The purpose of the present invention is to provide a depression auxiliary assessment method based on a large language model intelligent agent to address the deficiencies of the existing technology. It adopts a method of constructing a dynamic questioning intelligence, a multi-dimensional scoring intelligent agent and a clinical decision-making intelligent agent based on a large language model to realize the dynamic selection and adjustment of topic questions, match the corresponding scoring criteria based on the user's answer content, perform semantic understanding and emotional judgment on each question and answer, thereby completing the sub-item scoring of depressive symptoms, monitoring the conversation content and scoring results in real time, generating risk warnings and giving personalized diagnosis and intervention suggestions. This method can realize the structured digital expression of the HAM-D scale, construct a questioning mechanism with context understanding and strategy regulation capabilities, perform semantic scoring analysis on user responses, and integrate medical knowledge graphs to output clinical-level risk warnings and diagnosis and treatment suggestions, breaking through the limitations of traditional assessment models and improving the intelligent level of mental health screening and management. It realizes a dynamic interactive process that is closer to a real psychological interview, has context perception, adaptive guidance and semantic understanding capabilities, and has higher assessment accuracy, individual adaptability and clinical auxiliary value than traditional static questionnaire assessment methods.

[0006] The object of the present invention is achieved by providing a method for assisting depression assessment based on a large language model agent, wherein the method specifically comprises the following steps: Step 1: Structural processing is performed on the Hamilton Depression Rating Scale, which is converted into computer-readable JSON format data to facilitate subsequent information extraction and interaction design. Step 1 specifically includes: 1.1: Obtain the original Hamilton Depression Rating Scale text content including: scale item numbers, descriptive questions, scoring dimensions and corresponding scoring standards; 1.2: Perform semantic parsing and labeling on the above text content to extract structural elements including: entry topic, sentiment category, weight priority, and scoring logic; 1.3: Build a field mapping relationship based on a standardized template and convert the entry content into a JSON structure object containing fields such as "entry number", "problem description", "scoring criteria", "emotion label", and "clinical priority"; 1.4: Perform consistency check on the structured results to ensure that all entries have a unified format and clear semantics to meet the subsequent large language model input and call requirements; 1.5: Store the processed JSON format file in a database or intermediate data interface as the basic data resource for subsequent intelligent agent construction and question logic generation.

[0007] Step 2: Construct a dynamic questioning agent based on a large language model. Based on the clinical priority of the scale items and combined with the user's conversation context, dynamically select and adjust the topic questions. Step 2 specifically includes: 2.1: Based on the structured Hamilton Depression Rating Scale, extract the clinical priority, problem type, and scoring dimension of each item, and construct an item index table for scheduling; 2.2: Build a dynamic questioning agent driven by a large language model to receive user input and generate a representation of the current semantic context to guide the multi-round dialogue process; 2.3: Combine user conversation history and semantic context to dynamically select appropriate question items based on priority strategies, achieving personalized and risk-oriented query content. 2.4: Generate natural language question text for the selected item, and use a large language model to optimize the language and adjust the expression style to enhance the naturalness of the interaction; 2.5: User answers and question records are stored in a structured manner to provide standardized input data for subsequent scoring and decision-making modules.

[0008] Step 3: Construct a multi-dimensional scoring agent, match the corresponding scoring criteria based on the user's answer content, perform semantic understanding and emotional judgment on each question and answer, and thus complete the sub-item scoring of depressive symptoms. Step 3 specifically includes: 3.1: Build a scoring agent based on a large language model to perform semantic analysis on the natural language text of user responses and extract psychological characteristics such as emotions, attitudes, and cognition; 3.2: Match the extracted results with the scoring criteria of the corresponding items in the Hamilton Depression Rating Scale to identify their performance levels in different scale dimensions; 3.3: Based on the scoring criteria for each item, combined with the language sentiment and expression strength, the user's quantitative score for the item is automatically determined; 3.4: Supports aggregate calculation of the scoring results of multiple items under the same dimension and outputs multi-dimensional intermediate evaluation results; 3.5: The scoring results are stored in a structured manner to provide quantitative support for subsequent clinical decision-making.

[0009] Step 4: Integrate the Chinese medical knowledge graph to build a clinical decision-making agent. The agent monitors the conversation content and scoring results in real time, generates risk warnings, and provides personalized diagnosis, treatment, and intervention recommendations. Step 4 specifically includes: 4.1: Introducing the Chinese medical knowledge graph, integrating medical knowledge such as depression-related symptoms, diagnosis and treatment plans, risk factors, etc., to provide decision support for intelligent agents; 4.2: Associate and match the preliminary scores generated by the scoring agent with the clinical rules in the knowledge graph to identify symptom patterns and severity; 4.3: Introducing the Chinese medical knowledge graph, integrating medical knowledge such as depression-related symptoms, diagnosis and treatment plans, risk factors, etc., to provide decision support for the intelligent agent; 4.4: Based on semantic monitoring and scoring analysis results, risk warning mechanisms are automatically triggered, such as identifying emergency situations such as suicidal tendencies and severe depression; 4.5: Generate personalized treatment recommendations based on comprehensive analysis results, including recommended treatment directions, intervention strategies, or referral recommendations, for reference by clinical staff.

[0010] Step 5: After all items are scored, the scores of each item are automatically summarized to generate a total score, the degree of depression is determined according to the Hamilton Depression Rating Scale grading criteria, and corresponding diagnosis and treatment recommendations and clinical reference information are output. Step 5 specifically includes: 5.1: Automatically summarize the preliminary scores of each Hamilton Depression Rating Scale item, calculate the total score, and determine the degree of depression (such as mild, moderate, severe, etc.) according to the Hamilton Depression Rating Scale scoring criteria; 5.2: Based on the judgment results, an assessment report is automatically generated, outputting clinical reference information including depression level, detailed scores for each item, and personalized diagnosis and treatment recommendations.

[0011] Compared with the prior art, the present invention has the following beneficial technical effects and significant technical progress: 1) This invention introduces a question-asking agent driven by a large language model, which can dynamically generate and adjust questions based on the conversation context, realize real-time interaction in natural language, simulate the real interview process, and be closer to psychological counseling practice, thereby enhancing user experience and evaluation reliability and validity.

[0012] 2) This invention integrates Chinese medical knowledge graphs to construct a clinical decision-making intelligent agent, which can analyze user language behavior and scoring results in real time, identify high-risk signals (such as self-harm tendencies, severe depression, etc.), and output personalized diagnosis and treatment recommendations and intervention measures, providing intelligent assistance for clinical decision-making.

[0013] 3) This invention has the capabilities of context perception, adaptive guidance, and semantic understanding. Compared with traditional static questionnaire assessment methods, it has higher assessment accuracy, individual adaptability, and clinical auxiliary value. It breaks through the limitations of traditional assessment models and greatly improves the intelligence level of mental health screening and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Flowchart of the present invention; Figure 2 This is a schematic diagram of the process of Example 1. DETAILED DESCRIPTION

[0015] See Figure 1 , the present invention comprises the following steps: Step 1: Structural processing of the Hamilton Depression Rating Scale is performed, and it is converted into computer-readable JSON format data to facilitate subsequent information extraction and interaction design. The specific steps include the following: 1.1: Obtain the original Hamilton Depression Rating Scale text content, including scale item numbers, descriptive questions, scoring dimensions, and corresponding scoring criteria; 1.2: Perform semantic parsing and labeling on the above content to extract structural elements, including entry topics, sentiment categories, weight priorities, and scoring logic; 1.3: Build a field mapping relationship based on a standardized template and convert the entry content into a JSON structure object containing fields such as "entry number", "problem description", "scoring criteria", "emotional label", and "clinical priority"; 1.4: Perform consistency check on the structured results to ensure that all entries have a unified format and clear semantics to meet the subsequent large language model input and call requirements; 1.5: Store the processed JSON format file in a database or intermediate data interface as the basic data resource for subsequent intelligent agent construction and question logic generation.

[0016] Step 2: Construct a dynamic questioning agent based on a large language model. Based on the clinical priority of the scale items and combined with the user's conversation context, dynamically select and adjust the topic questions. Step 2 specifically includes: 2.1: Based on the structured Hamilton Depression Rating Scale, extract the clinical priority, problem type, and scoring dimension of each item, and construct an item index table for scheduling; 2.2: Build a dynamic questioning agent driven by a large language model to receive user input and generate a representation of the current semantic context to guide the multi-round dialogue process; 2.3: Combine user conversation history and semantic context to dynamically select appropriate question items based on priority strategies, achieving personalized and risk-oriented query content. 2.4: Generate natural language question text for the selected item, and use a large language model to optimize the language and adjust the expression style to enhance the naturalness of the interaction; 2.5: User answers and question records are stored in a structured manner to provide standardized input data for subsequent scoring and decision-making modules.

[0017] Step 3: Construct a multi-dimensional scoring agent, match the corresponding scoring criteria based on the user's answer content, perform semantic understanding and emotional judgment on each question and answer, and thus complete the sub-item scoring of depressive symptoms. Step 3 specifically includes: 3.1: Build a scoring agent based on a large language model to perform semantic analysis on the natural language text of user responses and extract psychological characteristics such as emotions, attitudes, and cognition; 3.2: Match the extracted results with the scoring criteria of the corresponding items in the Hamilton Depression Rating Scale to identify their performance levels in different scale dimensions; 3.3: Based on the scoring criteria for each item, combined with the language sentiment and expression strength, the user's quantitative score for the item is automatically determined; 3.4: Supports aggregate calculation of the scoring results of multiple items under the same dimension and outputs multi-dimensional intermediate evaluation results; 3.5: The scoring results are stored in a structured manner to provide quantitative support for subsequent clinical decision-making.

[0018] Step 4: Integrate the Chinese medical knowledge graph to build a clinical decision-making agent. The agent monitors the conversation content and scoring results in real time, generates risk warnings, and provides personalized diagnosis, treatment, and intervention recommendations. Step 4 specifically includes: 4.1: Introducing the Chinese medical knowledge graph, integrating medical knowledge such as depression-related symptoms, diagnosis and treatment plans, risk factors, etc., to provide decision support for intelligent agents; 4.2: Associate and match the preliminary scores generated by the scoring agent with the clinical rules in the knowledge graph to identify symptom patterns and severity; 4.3: Introducing the Chinese medical knowledge graph, integrating medical knowledge such as depression-related symptoms, diagnosis and treatment plans, risk factors, etc., to provide decision support for the intelligent agent; 4.4: Based on semantic monitoring and scoring analysis results, risk warning mechanisms are automatically triggered, such as identifying emergency situations such as suicidal tendencies and severe depression; 4.5: Generate personalized diagnosis and treatment recommendations based on the comprehensive analysis results, including recommended treatment directions, intervention strategies or referral recommendations, for reference by clinical staff.

[0019] Step 5: After all items are scored, the scores of each item are automatically summarized to generate a total score, the degree of depression is determined according to the Hamilton Depression Rating Scale grading criteria, and corresponding diagnosis and treatment recommendations and clinical reference information are output. Step 5 specifically includes: 5.1: Automatically summarize the preliminary scores of each Hamilton Depression Rating Scale item, calculate the total score, and determine the degree of depression (such as mild, moderate, severe, etc.) according to the Hamilton Depression Rating Scale scoring criteria; 5.2: Based on the judgment results, an assessment report is automatically generated, outputting clinical reference information including depression level, detailed scores for each item, and personalized diagnosis and treatment recommendations.

[0020] The present invention will be further described below with reference to specific embodiments and accompanying drawings. Example 1

[0021] See Figure 2 The depression auxiliary assessment system constructed using the present invention includes: a questioning agent, a multidimensional scoring agent, and a clinical decision-making agent. Users use the questioning agent to simulate a real interview process and achieve real-time interaction in natural language. The multidimensional scoring agent completes the sub-item scoring of depression symptoms based on the context of the user's answer. The clinical decision-making agent provides diagnosis and treatment recommendations based on the context of the answer. The specific process is as follows: 1) Convert the Hamilton Depression Rating Scale into computer-readable JSON format data, and then assign its content to the dynamic questioning agent and the multi-dimensional scoring agent respectively.

[0022] 2) Build a dynamic questioning agent based on a large language model. This agent dynamically selects and adjusts topic questions based on the clinical priority of scale items and the user's conversation context. This invention introduces a large language model-driven questioning agent that dynamically generates and adjusts questions based on the conversation context, enabling real-time interaction in natural language, simulating the actual interview process, and more closely resonating with psychological counseling practice, enhancing the user experience and the reliability and validity of the assessment.

[0023] 3) Construct a multi-dimensional scoring agent that matches the corresponding scoring criteria based on the user's answer content, performs semantic understanding and emotional judgment on each question and answer, and thus completes the sub-item scoring of depressive symptoms.

[0024] 4) Integrating the Chinese medical knowledge graph to build a clinical decision-making agent. This agent monitors conversation content and scoring results in real time, generates risk warnings, and provides personalized diagnosis, treatment, and intervention recommendations. This invention integrates the Chinese medical knowledge graph to build a clinical decision-making agent. This agent can analyze user language behavior and scoring results in real time, identify high-risk signals (such as self-harm tendencies and severe depression), and output personalized diagnosis, treatment, and intervention measures, providing intelligent assistance for clinical decision-making.

[0025] 5) After all items are scored, the scores of each item are automatically summarized to generate a total score, and the degree of depression is determined according to the Hamilton Depression Rating Scale grading criteria.

[0026] The above description is only a preferred embodiment of the present invention, and certain modifications may be made thereto within the scope defined by the claims of the present invention, but all modifications will fall within the scope of protection of the present invention.

Claims

1. A depression-assisted assessment method based on a large language model agent, characterized in that: The system uses a dynamic questioning agent to simulate real-time interaction with users, builds a multi-dimensional scoring agent and a clinical decision-making agent, and makes depression symptom scores and treatment recommendations based on the context of the user's answers. The depression auxiliary assessment specifically includes: Step 1: Structuralize the Hamilton Depression Rating Scale and convert it into computer-readable JSON format data; Step 2: Build a dynamic questioning agent based on a large language model. This agent dynamically selects and adjusts topic questions based on the clinical priority of Hamilton Depression Rating Scale items and the user conversation context. Step 3: Construct a multi-dimensional scoring agent. This agent matches the corresponding scoring criteria based on the context of the user's answer, performs semantic understanding and emotional judgment on each question and answer, and completes the sub-item scoring of depressive symptoms. Step 4: Integrate the Chinese medical knowledge graph to build a clinical decision-making agent. This agent monitors the conversation content and scoring results in real time, generates risk warnings, and provides personalized diagnosis, treatment, and intervention recommendations. Step 5: Automatically summarize the scores of each item to generate a total score, and use the Hamilton Depression Rating Scale grading standard to determine the degree of depression based on the generated total score, and output diagnosis and treatment recommendations and clinical reference information.

2. The depression auxiliary assessment method based on a large language model agent according to claim 1 is characterized in that: The step 1 is specifically as follows: 1.1: Obtain the original Hamilton Depression Rating Scale text content including scale item numbers, descriptive questions, scoring dimensions and corresponding scoring standards; 1.2: Perform semantic parsing and labeling on the above text content to extract structural elements including entry topics, sentiment categories, weight priorities, and scoring logic; 1.3: Build a field mapping relationship based on a standardized template and convert the entry content into a JSON structure object containing: entry number, problem description, scoring criteria, sentiment label, and clinical priority fields; 1.4: Perform consistency check on the structured results to ensure they are in a unified format. 1.5: Store the completed JSON format file in a database or intermediate data interface as the basic data resource for agent construction and question logic generation.

3. The depression auxiliary assessment method based on a large language model agent according to claim 1 is characterized in that: The step 2 is specifically as follows: 2.1: Based on the structured Hamilton Depression Rating Scale, extract the clinical priority, problem type, and scoring dimensions of each item, and construct an item index table for scheduling; 2.2: Build a dynamic questioning agent driven by a large language model to receive user input and generate a representation of the current semantic context; 2.3: Combine user conversation history and semantic context to dynamically select appropriate question items based on priority strategies, achieving personalized and risk-oriented query content. 2.4: Generate natural language question text for the selected item, and use a large language model to optimize the language and adjust the expression style to enhance the naturalness of the interaction; 2.5: Store user answers and question records in a structured manner.

4. The depression auxiliary assessment method based on a large language model agent according to claim 1 is characterized in that: The step 3 is specifically as follows: 3.1: Build a multi-dimensional scoring agent based on a large language model to perform semantic analysis on the natural language text of user responses and extract psychological characteristics including emotions, attitudes, and cognition; 3.2: Match the extracted psychological characteristics with the scoring criteria of the corresponding items in the Hamilton Depression Rating Scale to identify their performance levels in different scale dimensions; 3.3: The multi-dimensional scoring agent automatically determines the user's quantitative score for each item based on the scoring criteria of each item, combined with the language sentiment tendency and expression strength, and generates a preliminary score for the multi-dimensional scoring agent; 3.4: Aggregate the scoring results of multiple items under the same dimension and output the intermediate evaluation results of multiple dimensions; 3.5: Store the evaluation results in a structured manner.

5. The depression auxiliary assessment method based on a large language model agent according to claim 1 is characterized in that: The step 4 is specifically as follows: 4.1: Introducing the Chinese medical knowledge graph to integrate depression-related symptoms, treatment options, and risk factors; 4.2: Correlate and match the preliminary scores generated by the multi-dimensional scoring agent with the clinical rules in the knowledge graph to identify symptom patterns and severity; 4.3: Introducing the Chinese medical knowledge graph to integrate medical knowledge related to depression, including symptoms, diagnosis and treatment plans, risk factors, etc. 4.4: Automatically trigger risk warnings for suicidal tendencies or severe depression based on comprehensive analysis of semantic monitoring and assessment results; 4.5: Based on the comprehensive analysis results, automatically generate personalized diagnosis and treatment recommendations including recommended treatment directions, intervention strategies or referral suggestions.

6. The depression auxiliary assessment method based on a large language model agent according to claim 1 is characterized in that: The step 5 is specifically as follows: 5.1: The preliminary scores of each item of the Hamilton Depression Rating Scale are automatically summarized, and the degree of depression is determined by the Hamilton Depression Rating Scale scoring criteria based on the calculated total score, i.e. mild, moderate or severe; 5.2: Automatically generate an assessment report based on the depression level, and output clinical reference information including depression level, detailed scoring of each item, and personalized diagnosis and treatment recommendations.

Images