Depression risk screening optimization method and system based on large and small model linkage

Abstract

The invention discloses a depression risk screening optimization method based on large and small model linkage. The method comprises the following steps: selecting depression related indexes, obtaining interviewee questionnaire data, and preprocessing the data to obtain a scale source database; a depression risk prediction model is constructed, and PHQ-9 measurement results are compared for model training and verification; training a dialogue strategy module of a semantic analysis enhanced fine-tuning training large language model, constructing answer mapping through dynamic question generation and dialogue flow control, and converting a natural language of a user into standardized data required by a small model; training a man-machine interaction reinforcement learning model, generating a depression risk screening result based on a small model, inviting a user to carry out recognition degree evaluation, and dividing feedback into two types of recognition and question; for different feedbacks, strengthening or correcting the current interaction strategy and prediction logic, and storing the audited data as high-quality data to a training database by the system for subsequent large model fine tuning; and outputting a result and performing result interpretation and suggestion by using the semantic analysis reinforced fine-tuning large language model. Hierarchical early screening of depression risks is carried out based on large and small model linkage.

Description

Technical Field

[0001] The present invention belongs to the field of medical health technology, and in particular relates to a depression risk screening optimization method and system based on large and small model linkage. Background Art

[0002] Depression is a common mental illness characterized by persistent low mood, loss of interest or pleasure, and decreased energy. The prevalence of depression continues to rise worldwide. According to statistics from the World Health Organization (WHO), depression has become one of the leading causes of disability worldwide. Early screening and timely intervention are crucial for the effective treatment of depression. Currently, one of the most widely used depression risk screening tools in clinical practice is the Patient Health Questionnaire (PHQ-9). This scale uses a series of standardized questions to screen for the core symptoms of depression. Although this scale has certain clinical value in depression risk screening, it presents several limitations in the elderly population, including: a large number of negative leading questions, high respondent bias, neglect of somatization symptoms, and poor answering burden and experience.

[0003] With the advancement of machine learning and natural language processing (NLP) technologies, big data-based depression risk screening methods are gaining popularity. However, existing machine learning models still face technical difficulties. For one thing, feature engineering is limited by pre-defined question frameworks, resulting in poor ability to capture unstructured information features and reduced cross-group generalization. Furthermore, these systems still rely on passive data collection, failing to address the interactive user experience pain points of traditional questionnaires. While some models based on multimodal data can achieve seamless monitoring, further improvements are needed in terms of privacy protection and generalizability.

[0004] Although large language models such as GPT-4 and LLaMA have shown potential in simulating doctor-patient conversations, their technical flaws in depression screening have sparked academic skepticism: (1) semantic understanding lacks the constraints of a clinical theoretical framework; (2) generated content may contain medical factual errors; and (3) probabilistically generated conversation paths lead to impaired screening logic coherence or information omissions, especially as the number of conversation turns increases, the risk of missing key indicators also increases. More seriously, the "hallucination" phenomenon of large models means that most risk assessment conclusions lack a traceable chain of evidence to support them.

[0005] There are two key technical gaps in current research: (1) the disconnect between conversational interaction and theoretical frameworks: existing conversational systems only implement surface-level natural language interaction and fail to establish a multi-dimensional mapping mechanism with medical diagnostic standards; and (2) a technical fragmentation in risk stratification: large language models with over 10 billion hyperparameters still exhibit instability in five-class classification tasks, while traditional machine learning models struggle to process unstructured conversational data. This technical gap has led to an irreconcilable contradiction between user experience and medical reliability in existing systems.

[0006] In response to the above-mentioned technical deficiencies, the present invention constructs a large-scale model linkage architecture: the front end adopts a large language dialogue engine that has been fine-tuned in the domain, and realizes the collection of depression risk indicators through dynamic path planning, breaking the traditional questionnaire answering mode while improving the key indicator collection rate; the back end deploys a convolutional neural network model, based on the feature fusion of biological-psychological-social indicator data, to accurately realize multi-level depression risk screening. Therefore, the early screening of depression risk stratification based on the linkage of large and small models has important practical significance and broad application prospects. Summary of the Invention

[0007] The present invention aims to overcome the above-mentioned shortcomings of the existing technology and provide a depression risk screening optimization method and system based on the linkage of large and small models. It not only avoids negative guiding indicators in construction, but also uses the bio-psycho-social model as a theoretical framework for selecting screening prediction indicators in content architecture setting, and can also realize depression risk grading prediction based on natural language interaction.

[0008] The technical purpose of the present invention is achieved in this way:

[0009] The first aspect of the present invention is a depression risk screening optimization method based on large and small model linkage, comprising the following steps:

[0010] S1: Select depression-related indicators, obtain the respondents' questionnaire data, complete data preprocessing, and obtain the scale source database.

[0011] S2: Construct a depression risk prediction model and compare the PHQ-9 measurement results for model training and validation.

[0012] S3: Train semantic analysis to strengthen fine-tune the dialogue strategy module of the large language model. Through dynamic question generation and dialogue flow control, it builds answer mapping and converts user natural language into standardized data required by the small model.

[0013] S4: Train a reinforcement learning model for human-computer interaction. Based on the depression risk screening results generated by the small model, users are invited to evaluate their approval and the feedback is divided into two categories: "approval" and "questioning." For different feedback, the current interaction strategy and prediction logic are strengthened or revised. The system stores the reviewed data as high-quality data in the training database for subsequent fine-tuning of the large model.

[0014] S5: The semantic analysis is used to enhance the fine-tuning of the large language model to output the results and provide interpretation and suggestions.

[0015] Preferably, step S1 includes:

[0016] Based on literature research, we selected influencing factors related to depression, obtained relevant scales and questionnaire items for evaluating influencing factors, integrated them into an electronic comprehensive survey questionnaire, and collected data through field research.

[0017] Among them, in the biological aspect, the scale items used include personal disease history and family history assessment items (cardiovascular disease, metabolic disease, liver and kidney disease, etc.), physical symptom scale items (stomach pain, back pain, arm and leg and joint pain, headache, chest pain, dizziness, weakness, palpitations, shortness of breath, constipation or diarrhea or intestinal discomfort, fatigue, sleep problems), number of medications, exercise habits and intensity, sedentary habits, sleep status (duration of falling asleep and difficulty sleeping), and self-assessment of health status and management burden.

[0018] In terms of psychology, emotional state items and PHQ-9 scale items were used.

[0019] Among them, in the social aspect, the investigation includes the convenience of travel, procurement convenience, green exercise resources, community health service resources, and medical accessibility in the living environment; and the investigation includes the health-related economic burden and degree of social support in the economic and social aspects.

[0020] Furthermore, single data with more than 60% missing values were removed, and the remaining data were filled with missing values using multiple difference filling. Categorical variables used frequencies instead of original labels, and then the frequency values were Z-score standardized and outliers were eliminated; continuous variables were Z-score standardized and outliers were eliminated. The Z-score formula is:

[0021]

[0022] Where Z is the Z-Score value, X is a single data point in the dataset, μ is the mean of the dataset, and σ is the standard deviation of the dataset.

[0023] Furthermore, we preprocess the source database of the scale and use data augmentation based on statistical distribution as a dataset balancing method. Assuming that each feature follows a normal distribution (N(μ,σ^2)), we generate new data points through multiple sampling. The generated new samples (x^') satisfy:

[0024]

[0025] For categorical variables (y), the new samples generated by interpolation satisfy:

[0026] y new =argminy|P(y|X)-P(y|Xnew)| (3)

[0027] Among them, (P(y|X)) is the existing depression positive distribution, (P(y|X new )) is the generated depression-positive distribution.

[0028] Preferably, step S2 includes:

[0029] S21. Use convolutional neural network to build prediction model.

[0030] S22. Model training involves dividing the training set data into multiple batches and calculating and printing the training loss after each cycle.

[0031] S23. Model validation includes: calculating the accuracy, precision, recall, and F1 score of the prediction results, printing a detailed classification report, and plotting the confusion matrix to visualize the model's predictive performance.

[0032] Further, step S21 includes:

[0033] Correlation analysis, chi-square test and collinearity analysis were performed on each indicator in the scale source database and the depression risk results to complete the indicator dimensionality reduction.

[0034] Construct a convolutional neural network prediction model, which includes two convolutional layers, a LeakyReLU activation function, a Dropout layer, and a fully connected layer. The output size of the convolutional layer can be calculated based on the input size, convolution kernel size, and stride, as shown in the following formula:

[0035]

[0036] Among them, L in is the input size, k is the convolution kernel size, and s is the step size.

[0037] For the calculation of the output size of the two convolutional layers, the formula is as follows:

[0038] L out =(L in-k1+1)-k2+1 (5)

[0039] Among them, k1 and k2 are the sizes of the first and second layer convolution kernels respectively.

[0040] The Dropout layer randomly drops some neurons to prevent overfitting. The dropout probability is p, and the retention probability is 1-p. During the training process, the output x of each neuron is set to zero with probability p and remains unchanged with probability 1-p. The formula is as follows:

[0041]

[0042] The output y of the fully connected layer can be expressed as the product of the input feature x and the weight matrix W plus the bias b, as follows:

[0043] y=Wx+b (7)

[0044] Further, step S22 includes:

[0045] The training set data is divided into multiple batches for model training. After each cycle, the training loss is calculated and printed out.

[0046] The CNN model is trained using the Adam optimizer, and the cross-entropy loss function is used to measure the difference between the model's predicted value ypred and the true label ytrue:

[0047]

[0048] Where ytrue,i is the one-hot encoding of the true label of the depression risk state, and ypred,i is the probability distribution of the depression risk prediction.

[0049] Further, step S23 includes:

[0050] Validate the model on the validation set, make predictions for the trained model, calculate the accuracy, precision, recall, and F1 score of the prediction results, and print a detailed classification report. Draw the confusion matrix C as follows:

[0051] The confusion matrix C is a K×K matrix, where K is the number of categories. The element Cij in the i-th row and j-th column of the matrix represents the number of samples that are actually in depression risk category i but are predicted to be in depression risk category j:

[0052]

[0053] Preferably, step S3 includes fine-tuning of the large language model, interactive design, and answer mapping mechanism.

[0054] The fine-tuning and interaction design of the large language model includes three stages: instruction fine-tuning, reinforcement learning optimization, and consistency constraint fine-tuning. The three stages are executed sequentially and fine-tuning is performed on parameters at different levels.

[0055] The instruction fine-tuning phase uses a fine-tuning dataset primarily based on structured question-answering generated by expert annotations and templates to inject domain knowledge. This includes dataset acquisition, dynamic weight adjustment, and domain knowledge injection techniques.

[0056] Dataset acquisition involves collecting high-quality corpus from the mental health field, annotating it with expert annotations based on keyword matching, sentiment intensity, question format, and question path, to create a domain-specific fine-tuning dataset. Public literature databases are manually integrated to form an external medical database, and instruction templates are designed for supervised training.

[0057] Dynamic weight adjustment includes: on this basis, adjusting the loss function weights of different samples according to domain relevance and task complexity.

[0058] Domain knowledge injection includes: using external medical knowledge bases (such as DSM-5 diagnostic criteria) to post-process and enhance the model.

[0059] Specifically include:

[0060] Design a weighted loss function L w , by assigning different weights w to different samples i To optimize model performance. The weighted loss function formula is as follows:

[0061]

[0062] Where N is the total number of samples, w i represents the weight of the i-th sample, which is determined by the depression risk correlation or depression risk level, L(y i ,y i ) is the loss function for a single sample.

[0063] Weight w i The calculation can be based on the importance score s of the sample i :

[0064]

[0065] Among them, s i It is calculated through psychiatric expert annotation and depression risk judgment complexity indicators (such as question difficulty and user answer ambiguity).

[0066] The reinforcement learning optimization phase uses real multi-round conversation data for fine-tuning training, aiming to enable the model to generate responses that better meet user expectations, specifically the next question generation. This phase primarily includes a context encoder, an intent recognition module, dialogue state tracking (DST), and a reinforcement learning optimization strategy.

[0067] The context encoder uses the Transformer architecture built into the large language model to directly jointly model the user's historical conversations and current input.

[0068] The intent recognition model classifies the intent of user input through pre-training, supports fuzzy matching and multi-intent parsing, and understands the emotional tendencies in user answers (such as insomnia and loss of interest).

[0069] The dialogue state tracking module is based on a memory network and dynamically updates the user's answer status, including covered and uncovered emotional topics.

[0070] The reinforcement learning optimization strategy is to introduce the reinforcement learning algorithm (PPO) to stop dialogue generation when the model completes the collection of standardized data features within a limited number of rounds. The goal of PPO is to minimize the agent objective function L CLIP :

[0071] L CLIP (θ)=E t [min(r t (θ)A t ,clip(r t (θ),1-ε,1+ε)A t )] (12)

[0072] in, A represents the probability ratio of the trajectory of the new generated problem to the trajectory of the old problem, t It is the advantage function, which indicates the advantage of the current question generation over the average normative question answering. The value of ε is usually between 0.1 and 0.2, indicating the clipping range. clip(x,a,b) means clipping x to the interval [a,b].

[0073] Advantage function A t The formula is as follows:

[0074] V(s t )A t =Q(s t ,a t )-V(s t ) (13)

[0075] Among them, Q(s t ,a t ) represents the state-action value function, V(s t) represents the state-value function.

[0076] Consistency constraint fine-tuning is the generation consistency control driven by the attention mechanism, which includes global attention constraints and local attention constraints.

[0077] Specifically include:

[0078] Global attention constraint: A global attention mechanism is introduced during the generation process to force the model to focus on the entire conversation history and questionnaire structure, ensuring that the generated content is consistent with the context.

[0079] Consistency loss function: A consistency loss function based on contrastive learning is designed to penalize the semantic deviation between the generated content and the reference answer, thereby improving the accuracy of the generated results.

[0080]

[0081] Among them, x i represents the anchor point sample, represents a positive sample (consistent with the anchor point), represents a negative sample (not consistent with the anchor point), d(x,y) represents the distance metric between samples (such as cosine similarity or Euclidean distance), y i represents the sample label (1 represents a positive sample, 0 represents a negative sample), and m is the boundary threshold.

[0082] Local attention constraints: When generating each question, local constraint modules (such as template matching and grammar checking) are added to ensure that the generated content conforms to language specifications and logical requirements.

[0083] Autoregressive verification: After the content is generated and before it is delivered to the user, the generated content is retrospectively analyzed using an autoregressive verification mechanism by simulating multiple rounds of dialogue to detect potential logical conflicts or inconsistencies and automatically correct them.

[0084] Autoregressive prediction formula:

[0085]

[0086] Among them, x t Indicates that interaction with the user is the tth generated question, x ＜t represents the first t-1 generated questions.

[0087] When checking for logical conflicts, you can calculate the KL divergence between the generated content and the reference question:

[0088]

[0089] Here, P(x) represents the distribution of generated content, and Q(x) represents the distribution of original reference questions.

[0090] Among them, the answer mapping mechanism adopts a context encoder to extract the emotional intensity level in the conversation, align the keyword matching with the questionnaire options, and build an answer mapping rule library based on rules and statistical methods.

[0091] Preferably, in step S4, after outputting the user's hierarchical outcomes and label annotations, the interactive interface sets an "approval" or "questioning" channel.

[0092] When the user feedback is "approval", the system marks it as "high-quality data" and stores it in the training database.

[0093] When user feedback is "questionable", the interactive content of the "question" will be marked as "data to be verified", triggering the model's self-correction mechanism and mining a better path from historical data. Detailed feedback (including specific question points and improvement suggestions) will be recorded and stored in the training database after manual review or expert annotation, and updated to "high-quality data".

[0094] By analyzing feedback data, key interaction indicators (such as emotional expression patterns, lifestyle descriptions, etc.) are extracted and incorporated into feature engineering of the prediction model.

[0095] Through regular statistical analysis, we can identify high-frequency problems or common demands, and guide the optimization of prediction logic and interaction methods.

[0096] The aforementioned reinforcement learning of “high-quality data” for subsequent model fine-tuning includes:

[0097] By giving positive reward signals, strengthening the current interaction strategy and prediction logic, and using high-quality data for incremental reinforcement learning training, the environment is defined: the state is the current conversation history, including all questions raised by the system and the user's answers, represented by a sequence:

[0098] S t ={(q1,a1),(q2,a2),…,(q t ,a t )} (17)

[0099] where q i is the i-th question, a i is the user’s answer. Action is the next question generated by the system based on the current depression risk screening status. t+1 Reward is the user’s recognition of the final evaluation result.

[0100] Use the optimization strategy network π θ (a t |s t ), that is, in a given state s tNext select action a t The probability distribution of maximizing the long-term cumulative reward. In each complete round of user research, the following conversation history is recorded:

[0101] S t ={(q1,a1),(q2,a2),…,(q t ,a t )} (18)

[0102] Evaluation result y and user approval label R. The reward function R(ST) is defined as:

[0103]

[0104] Use high-quality data to update the strategy. For each high-quality dialogue logic trajectory τ, calculate the gradient and update the parameters, where α is the learning rate:

[0105]

[0106] A second aspect of the present invention relates to a depression risk screening optimization system based on large and small model linkage, which is used to implement a depression risk screening optimization method based on large and small model linkage of the present invention, and the system includes:

[0107] The data collection end, which includes the scale source database, fine-tuning dataset, external medical database and training dataset, is used to collect and save data information and upload the data information to the cloud server;

[0108] The cloud server is deployed with a data analysis module, a trained convolutional neural network model, a trained semantic analysis enhanced fine-tuning large language model and a trained human-computer interaction reinforcement learning model. The data analysis module is equipped with data missing value interpolation, standardization processing, data format unification, data basic statistical analysis and feature engineering algorithms. The cloud server is used to pre-process the collected digital and text data, call the trained semantic analysis enhanced fine-tuning large language model for dialogue generation and answer mapping, and further call the trained convolutional neural network model to predict the user's depression risk level.

[0109] Feedback module, used to provide feedback to users on their evaluation of prediction results;

[0110] A display module, used to display the depression risk level prediction results;

[0111] The early warning module is used to remind users, provide explanatory explanations based on the depression risk level prediction results, and provide relief suggestions.

[0112] The present invention has the following beneficial effects:

[0113] (1) The present invention avoids secondary trauma or negative impact during the screening process of patients at risk of depression by avoiding negative guidance indicators, thus helping to ensure patient safety;

[0114] (2) The present invention adopts the bio-psycho-social model as the modeling framework, which implements the medical theoretical framework guidance of large and small models, and contributes to the scientificity and interpretability of the predicted outcomes;

[0115] (3) The present invention uses convolutional neural network technology to achieve accurate prediction of depression risk stratification, which has a positive impact on early screening and intervention treatment of high-risk depression populations;

[0116] (4) This invention uses large language model fine-tuning and interaction design technology, and through open-ended questions and optimal dialogue logic trajectory reinforcement learning, it achieves the transformation from traditional stereotyped answering to natural communication screening mode, thus improving the interactive experience;

[0117] (5) The present invention adopts the technology of linking large and small models, uses a machine learning model to achieve depression risk classification prediction, and uses a large language model to improve data collection and effect feedback, which helps to accurately screen and improve the feasibility of home application.

[0118] (6) The present invention adopts the reinforcement learning model technology of human-computer interaction to realize the screening and optimization of large and small models for real-world application feedback, which helps to maintain the adaptability of the linkage between large and small models in long-term applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0119] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. The drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0120] Figure 1 It is the principle framework diagram of the method of the present invention;

[0121] Figure 2 is a flow chart of the method of the present invention;

[0122] Figure 3 It is a technical implementation diagram of the present invention;

[0123] Figure 4 This is a framework diagram of the convolutional neural network prediction model of the present invention;

[0124] Figure 5 The framework diagram of the semantic analysis enhancement and fine-tuning large language model of the present invention;

[0125] Figure 6This is a flow chart of the human-computer interaction reinforcement learning model of the present invention;

[0126] Figure 7 This is a system structure diagram of the present invention. Specific implementation methods

[0127] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0128] Example 1

[0129] like Figure 2 As shown, the present invention provides a method for identifying depression by combining a precise model with a lightweight model, comprising the following steps:

[0130] S1: Select depression-related indicators, obtain the respondents' questionnaire data, complete data preprocessing, and obtain the scale source database.

[0131] S2: Construct a depression risk prediction model and compare the PHQ-9 measurement results for model training and validation.

[0132] S3: Train semantic analysis to enhance fine-tune the large language model, and form complete standardized data collection through dynamic question generation and answer mapping.

[0133] S4: Train a reinforcement learning model for human-computer interaction. The system generates depression risk screening results and invites users to rate their approval, categorizing feedback as either "approval" or "questioning." Based on different feedback, the system strengthens or modifies the current interaction strategy and prediction logic. The system then stores this high-quality data in a training database for subsequent model fine-tuning.

[0134] S5: The semantic analysis is used to enhance the fine-tuning of the large language model to output the results and provide interpretation and suggestions.

[0135] See also Figure 3The specific implementation process of the present invention is as follows: the investigator uses a mobile phone or computer to collect questionnaire data, performs preprocessing and processes the results according to the scoring method of the questionnaire and the scale itself to obtain the scale source database, performs data balancing on the scale source database based on statistical distribution data enhancement, and trains the convolutional neural network model; based on semantic analysis and enhanced fine-tuning of the large language, user data is collected through dialogue generation and answer mapping, and the data is input into the convolutional neural network model for prediction to determine the user's depression risk stratification. If the user feedback is "questioning", the PHQ-9 scale is output for standardized evaluation and the screening ends; if the user feedback is "approving", the screening ends.

[0136] Specifically, step S1 includes:

[0137] like Figure 1 As shown in the figure, based on literature research, influencing factors related to depression were selected, relevant scales and questionnaire items for evaluating influencing factors were obtained, integrated into an electronic comprehensive survey questionnaire, and data were collected through field research.

[0138] Among them, in the biological aspect, the scale items used include personal disease history and family history assessment items (cardiovascular disease, metabolic disease, liver and kidney disease, etc.), physical symptom scale items (stomach pain, back pain, arm and leg and joint pain, headache, chest pain, dizziness, weakness, palpitations, shortness of breath, constipation or diarrhea or intestinal discomfort, fatigue, sleep problems), number of medications, exercise habits and intensity, sedentary habits, sleep status (duration of falling asleep and difficulty sleeping), and self-assessment of health status and management burden.

[0139] In terms of psychology, emotional state items and PHQ-9 scale items were used.

[0140] Among them, in the social aspect, the investigation includes the convenience of travel, procurement convenience, green exercise resources, community health service resources, and medical accessibility in the living environment; and the investigation includes the health-related economic burden and degree of social support in the economic and social aspects.

[0141] Furthermore, single data with more than 60% missing values were removed, and the remaining data were filled with missing values using multiple difference filling. Categorical variables used frequencies instead of original labels, and then the frequency values were Z-score standardized and outliers were eliminated; continuous variables were Z-score standardized and outliers were eliminated. The Z-score formula is:

[0142]

[0143] Where Z is the Z-Score value, X is a single data point in the dataset, μ is the mean of the dataset, and σ is the standard deviation of the dataset.

[0144] Furthermore, we preprocess the source database of the scale and use data augmentation based on statistical distribution as a dataset balancing method. Assuming that each feature follows a normal distribution (N(μ,σ^2)), we generate new data points through multiple sampling. The generated new samples (x^') satisfy:

[0145]

[0146] For categorical variables (y), the new samples generated by interpolation satisfy:

[0147] y new =argminy|P(y|X)-P(y|Xnew)| (3)

[0148] Among them, P(y|X) is the existing depression-positive distribution, and P(y|Xnew) is the generated depression-positive distribution.

[0149] Step S1 is described in detail with a specific example.

[0150] In terms of bio-psycho-social aspects, the indicators are first distinguished between categorical variables and continuous variables.

[0151] Furthermore, if more than 60% of a patient's overall data was missing, the patient was excluded. For Patient A, 5% of the data in the questionnaire was missing; therefore, the patient was included, and the remaining data were imputed using multiple interpolation. For categorical variables, the original labels were replaced with frequencies, and the frequencies were then Z-score normalized, with outliers removed. Continuous variables were also Z-score normalized, with outliers removed.

[0152] Furthermore, in the data augmentation part, the continuous variable is the depression screening score of PHQ-9. The depression-positive group data in the dataset follows the normal distribution N(18, 3), with an average PHQ-9 score of 18 and a standard deviation of 3. New data points are generated through multiple sampling to balance the number of samples in the dataset. The new sample formula satisfies:

[0153] y new =arg m y |P(y|X)-P(y|X new )| (3)

[0154] Among them, P(y|X) is the distribution of existing depression-positive questions, P(y|X new ) is the distribution of generated depression-positive questions.

[0155] Taking the existence of difficulty falling asleep as an example, the categorical variable uses the interpolation method to generate new samples. Given the feature X, the posterior probability P(y|X) of y is maximized, y newWill be used as the label of the newly generated sample, the new sample satisfies:

[0156] y new =argmax y P(y|X) (21)

[0157] like Figure 4 , build a convolutional neural network model, which includes indicator dimensionality reduction, input layer, two convolutional layers, a Dropout layer, a fully connected layer and an output layer. The training process is as follows:

[0158] Correlation analysis, chi-square test and collinearity analysis were performed on each indicator in the scale source database and the depression risk results to complete the indicator dimensionality reduction.

[0159] Construct a convolutional neural network prediction model, which includes two convolutional layers, a LeakyReLU activation function, a Dropout layer, and a fully connected layer. The output size of the convolutional layer can be calculated based on the input size, convolution kernel size, and stride, as shown in the following formula:

[0160]

[0161] Among them, L in is the input size, k is the convolution kernel size, and s is the step size.

[0162] For the calculation of the output size of the two convolutional layers, the formula is as follows:

[0163] L out =(L in -k1+1)-k2+1 (6)

[0164] Among them, k1 and k2 are the sizes of the first and second layer convolution kernels respectively.

[0165] The Dropout layer randomly drops some neurons to prevent overfitting. The dropout probability is p, and the retention probability is 1-p. During the training process, the output x of each neuron is set to zero with probability p and remains unchanged with probability 1-p. The formula is as follows:

[0166]

[0167] The output y of the fully connected layer can be expressed as the product of the input feature x and the weight matrix W plus the bias b, as follows:

[0168] y=Wx+b (7)

[0169] The training set data is divided into multiple batches for model training. After each cycle, the training loss is calculated and printed out.

[0170] The CNN model is trained using the Adam optimizer, and the cross-entropy loss function is used to measure the difference between the model's predicted value ypred and the true label ytrue:

[0171]

[0172] Where ytrue,i is the one-hot encoding of the true label of the depression risk state, and ypred,i is the probability distribution of the depression risk prediction.

[0173] Validate the model on the validation set, make predictions for the trained model, calculate the accuracy, precision, recall, and F1 score of the prediction results, and print a detailed classification report. Draw the confusion matrix C as follows:

[0174] The confusion matrix C is a K×K matrix, where K is the number of categories. The element Cij in the i-th row and j-th column of the matrix represents the number of samples that are actually in depression risk category i but are predicted to be in depression risk category j:

[0175]

[0176] like Figure 5 The fine-tuning and interaction design of the large language model and the answer mapping mechanism framework include instruction fine-tuning, reinforcement learning optimization, and consistency constraint fine-tuning. The three stages are executed sequentially and fine-tuning is performed on parameters at different levels. The answer mapping mechanism includes a context encoder and a statistical method-based answer mapping rule base. The training process is as follows:

[0177] First, complete the fine-tuning and interaction design of the large language model, sequentially perform instruction fine-tuning, reinforcement learning optimization, and consistency constraint fine-tuning, and fine-tune parameters at different levels.

[0178] The instruction fine-tuning phase uses a fine-tuning dataset primarily based on structured question-answering generated by expert annotations and templates to inject domain knowledge. This includes dataset acquisition, dynamic weight adjustment, and domain knowledge injection techniques.

[0179] Dataset acquisition involves collecting high-quality corpus from the mental health field, annotating it with expert annotations based on keyword matching, sentiment intensity, question format, and question path, to create a domain-specific fine-tuning dataset. Public literature databases are manually integrated to form an external medical database, and instruction templates are designed for supervised training.

[0180] Dynamic weight adjustment includes: on this basis, adjusting the loss function weights of different samples according to domain relevance and task complexity.

[0181] Domain knowledge injection includes: using external medical knowledge bases (such as DSM-5 diagnostic criteria) to post-process and enhance the model.

[0182] Specifically include:

[0183] Design a weighted loss function L w , by assigning different weights w to different samples i To optimize model performance. The weighted loss function formula is as follows:

[0184]

[0185] Where N is the total number of samples, w i represents the weight of the i-th sample, which is determined by the depression risk correlation or depression risk level, L(y i ,y i ) is the loss function for a single sample.

[0186] Weight w i The calculation can be based on the importance score s of the sample i :

[0187]

[0188] Among them, s i It is calculated through psychiatric expert annotation and depression risk judgment complexity indicators (such as question difficulty and user answer ambiguity).

[0189] The reinforcement learning optimization phase uses real multi-round conversation data for fine-tuning training, aiming to enable the model to generate responses that better meet user expectations, specifically the next question generation. This phase primarily includes a context encoder, an intent recognition module, dialogue state tracking (DST), and a reinforcement learning optimization strategy.

[0190] The context encoder uses the Transformer architecture built into the large language model to directly jointly model the user's historical conversations and current input.

[0191] The intent recognition model classifies the intent of user input through pre-training, supports fuzzy matching and multi-intent parsing, and understands the emotional tendencies in user answers (such as insomnia and loss of interest).

[0192] The dialogue state tracking module is based on a memory network and dynamically updates the user's answer status, including covered and uncovered emotional topics.

[0193] The reinforcement learning optimization strategy is to introduce the reinforcement learning algorithm (PPO) to stop dialogue generation when the model completes the collection of standardized data features within a limited number of rounds. The goal of PPO is to minimize the agent objective function L CLIP :

[0194] L CLIP (θ)=E t [min(r t (θ)A t ,clip(r t (θ),1-ε,1+ε)A t )] (12)

[0195] in, A represents the probability ratio of the trajectory of the new generated problem to the trajectory of the old problem, t It is the advantage function, which indicates the advantage of the current question generation over the average normative question answering. The value of ε is usually between 0.1 and 0.2, indicating the clipping range. clip(x,a,b) means clipping x to the interval [a,b].

[0196] Advantage function A t The formula is as follows:

[0197] V(s t )A t =Q(s t ,a t )-V(s t ) (13)

[0198] Among them, Q(s t ,a t ) represents the state-action value function, V(s t ) represents the state-value function.

[0199] Consistency constraint fine-tuning is the generation consistency control driven by the attention mechanism, which includes global attention constraints and local attention constraints.

[0200] Specifically include:

[0201] Global attention constraint: A global attention mechanism is introduced during the generation process to force the model to focus on the entire conversation history and questionnaire structure, ensuring that the generated content is consistent with the context.

[0202] Consistency loss function: A consistency loss function based on contrastive learning is designed to penalize the semantic deviation between the generated content and the reference answer, thereby improving the accuracy of the generated results.

[0203]

[0204] Among them, x i represents the anchor point sample, represents a positive sample (consistent with the anchor point), represents a negative sample (not consistent with the anchor point), d(x,y) represents the distance metric between samples (such as cosine similarity or Euclidean distance), yi represents the sample label (1 represents a positive sample, 0 represents a negative sample), and m is the boundary threshold.

[0205] Local attention constraints: When generating each question, local constraint modules (such as template matching and grammar checking) are added to ensure that the generated content conforms to language specifications and logical requirements.

[0206] Autoregressive verification: After the content is generated and before it is delivered to the user, the generated content is retrospectively analyzed using an autoregressive verification mechanism by simulating multiple rounds of dialogue to detect potential logical conflicts or inconsistencies and automatically correct them.

[0207] Autoregressive prediction formula:

[0208]

[0209] Among them, x t Indicates that interaction with the user is the tth generated question, x ＜t represents the first t-1 generated questions.

[0210] When checking for logical conflicts, you can calculate the KL divergence between the generated content and the reference question:

[0211]

[0212] Here, P(x) represents the distribution of generated content, and Q(x) represents the distribution of original reference questions.

[0213] Furthermore, the answer mapping mechanism adopts a context encoder-based approach to extract emotion intensity levels in conversations, align keyword matching with questionnaire options, and construct an answer mapping rule library based on rules and statistical methods.

[0214] like Figure 6 The human-computer interaction reinforcement learning model consists of a user interaction layer, a data processing layer, a reinforcement learning optimization layer, and an output layer. The user interaction layer includes two feedback channels: user input and "approval" or "questioning." For the "questioning" channel, historical conversations are input into the data processing layer for self-revision, manual review, and feature engineering analysis. After data processing, the historical conversations from the "approval" channel are marked as "high-quality data" and enter the reinforcement learning optimization layer. They then sequentially enter the data layer, environment layer, and optimization layer to complete the training data set storage, environment layer settings, and optimization strategy network learning. The training process is as follows:

[0215] When the user feedback is "approval", the system marks it as "high-quality data" and stores it in the training database.

[0216] When user feedback is "questionable", the interactive content of the "question" will be marked as "data to be verified", triggering the model's self-correction mechanism and mining a better path from historical data. Detailed feedback (including specific question points and improvement suggestions) will be recorded and stored in the training database after manual review or expert annotation, and updated to "high-quality data".

[0217] By analyzing feedback data, key interaction indicators (such as emotional expression patterns, lifestyle descriptions, etc.) are extracted and incorporated into feature engineering of the prediction model.

[0218] Through regular statistical analysis, we can identify high-frequency problems or common demands, and guide the optimization of prediction logic and interaction methods.

[0219] Furthermore, reinforcement learning is performed on "high-quality data" for subsequent model fine-tuning, including:

[0220] By giving positive reward signals, strengthening the current interaction strategy and prediction logic, and using high-quality data for incremental reinforcement learning training, the environment is defined: the state is the current conversation history, including all questions raised by the system and the user's answers, represented by a sequence:

[0221] S t ={(q1,a1),(q2,a2),…,(q t ,a t )} (18)

[0222] where q i is the i-th question, a i is the user’s answer. Action is the next question generated by the system based on the current depression risk screening status. t+1 Reward is the user’s recognition of the final evaluation result.

[0223] Use the optimization strategy network π θ (a t |s t ), that is, in a given state s t Next select action a t The probability distribution of maximizing the long-term cumulative reward. In each complete round of user research, the following information is recorded:

[0224] Conversation History t ={(q1,a1),(q2,a2),…,(q t ,a t )};

[0225] Evaluation result y and user approval label R. The reward function R(ST) is defined as:

[0226]

[0227] Use high-quality data to update the strategy. For each high-quality dialogue logic trajectory τ, calculate the gradient and update the parameters, where α is the learning rate:

[0228]

[0229] Example 2

[0230] In addition, if Figure 7 As shown, this embodiment provides a depression risk screening optimization system based on large and small model linkage, which is used to implement the method described in Example 1. The system includes:

[0231] The data collection end, which includes the scale source database, fine-tuning dataset, external medical database and training dataset, is used to collect and save data information and upload the data information to the cloud server;

[0232] The cloud server is deployed with a data analysis module, a trained convolutional neural network model, a trained semantic analysis enhanced fine-tuning large language model and a trained human-computer interaction reinforcement learning model. The data analysis module is equipped with data missing value interpolation, standardization processing, data format unification, data basic statistical analysis and feature engineering algorithms. The cloud server is used to pre-process the collected digital and text data, call the trained semantic analysis enhanced fine-tuning large language model for dialogue generation and answer mapping, and further call the trained convolutional neural network model to predict the user's depression risk level.

[0233] Feedback module, used to provide feedback to users on their evaluation of prediction results;

[0234] A display module, used to display the depression risk level prediction results;

[0235] The early warning module is used to remind users, provide explanatory explanations based on the depression risk level prediction results, and provide relief suggestions.

[0236] When users use this system to screen their depression risk level, they only need to complete an interactive conversation with the large language model for about 5 minutes. The conversation data is automatically preprocessed, and semantic analysis and answer mapping are completed. The convolutional neural network model quickly predicts the depression risk level, displays the screening results on the display module, reminds the user through the early warning module, and provides suggestions for alleviating negative emotions.

[0237] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A depression risk screening optimization method based on large and small model linkage, characterized by The steps include: S1: Select depression-related indicators, obtain questionnaire data from respondents, complete data preprocessing, and obtain the scale source database; S2: Construct a depression risk prediction model and compare the PHQ-9 measurement results for model training and validation; S3: A dialogue strategy module that fine-tunes and trains a large language model based on semantic analysis, enabling dynamic question generation and dialogue flow control. It also builds an answer mapper to convert user natural language responses into standardized metrics required by the small model. S4: Train a reinforcement learning model for human-computer interaction. Based on the depression risk screening results generated by the small model, users are invited to rate their approval and categorize the feedback into "approval" and "questioning"; For different feedback, the system strengthens or amends the current interaction strategy and prediction logic. The system stores the reviewed data as high-quality data in the training database for subsequent fine-tuning of the large model. S5: The semantic analysis is used to enhance the fine-tuning of the large language model to output the results and provide interpretation and suggestions.

2. The depression risk screening optimization method based on large and small model linkage according to claim 1, characterized in that: Step S1 includes: Based on literature research, we selected influencing factors related to depression, obtained relevant scales and questionnaire items for evaluating influencing factors, integrated them into an electronic comprehensive survey questionnaire, and collected data through field research. Among them, in terms of biology, the scale items used include personal disease history and family history assessment items (cardiovascular disease, metabolic disease, liver and kidney disease, etc.), physical symptom scale items (stomach pain, back pain, arm and leg and joint pain, headache, chest pain, dizziness, weakness, palpitations, shortness of breath, constipation or diarrhea or intestinal discomfort, fatigue, sleep problems), number of medications, exercise habits and intensity, sedentary habits, sleep status (sleep time and sleep difficulties), and self-assessment of health status and management burden; Among them, in the psychological aspect, emotional state items and PHQ-9 scale items were used; In terms of society, the survey included travel convenience, purchasing convenience, greening and exercise resources, community health service resources, and medical accessibility in the living environment; and the survey also included health-related economic burdens and social support levels in the economic and social aspects. Furthermore, single data with more than 60% missing values were removed, and the remaining data were filled with missing values using multiple difference filling. Categorical variables used frequencies instead of original labels, and then the frequency values were Z-score standardized and outliers were eliminated; continuous variables were Z-score standardized and outliers were eliminated. The Z-score formula is: Where Z is the Z-Score value, X is a single data point in the data set, μ is the mean of the data set, and σ is the standard deviation of the data set; Furthermore, the scale source database is preprocessed, and data enhancement based on statistical distribution is used as a data set balancing method. Assuming that each feature obeys a normal distribution (N(μ,σ^2)), new data points are generated through multiple sampling, and the generated new samples (x^') satisfy: For categorical variables (y), the new samples generated by interpolation satisfy: y new =argminy|P(y|X)-P(y|Xnew)| (3) Where (P(y|X)) is the existing depression-positive distribution, (P(y|X new )) is the generated depression-positive distribution.

3. The depression risk screening optimization method based on large and small model linkage according to claim 1, characterized in that: Step S2 includes: S21. Use convolutional neural networks to build prediction models; S22. Model training involves dividing the training set data into batches and calculating and printing the training loss after each epoch. S23. Model validation includes calculating the accuracy, precision, recall, and F1 score of the prediction results, printing a detailed classification report, and plotting the confusion matrix to visualize the model's predictive performance.

4. The depression risk screening optimization method based on large and small model linkage according to claim 3, characterized in that: Step S21 includes: Correlation analysis, chi-square test, and collinearity analysis were performed on each indicator in the scale source database and the depression risk results to complete the indicator dimensionality reduction; Construct a convolutional neural network prediction model, which includes two convolutional layers, a LeakyReLU activation function, a Dropout layer, and a fully connected layer. The output size of the convolutional layer can be calculated based on the input size, convolution kernel size, and stride. The formula is as follows: Among them, L in is the input size, k is the convolution kernel size, and s is the step size; For the calculation of the output size of the two convolutional layers, the formula is as follows: L out =(L in -k1+1)-k2+1 (5) where k1 and k2 are the sizes of the convolution kernels of the first and second layers respectively; The Dropout layer randomly discards some neurons to prevent overfitting; the discard probability is p, and the retention probability is 1-p; during the training process, the output x of each neuron is set to zero with probability p and remains unchanged with probability 1-p. The formula is as follows: The output y of the fully connected layer can be expressed as the product of the input feature x and the weight matrix W plus the bias b, as follows: y=Wx+b (7) Step S22 includes: Divide the training set data into multiple batches for model training. After each cycle, calculate the training loss and print it out. The CNN model is trained using the Adam optimizer, and the cross-entropy loss function is used to measure the difference between the model's predicted value ypred and the true label ytrue: Where ytrue,i is the one-hot encoding of the true label of depression risk status, and ypred,i is the probability distribution of depression risk prediction; Step S23 includes: Validate the model on the validation set, make predictions for the trained model, calculate the accuracy, precision, recall, and F1 score of the prediction results, and print a detailed classification report; draw the confusion matrix C, as follows: The confusion matrix C is a K×K matrix, where K is the number of categories; the element Cij in the i-th row and j-th column of the matrix represents the number of samples that are actually in depression risk category i but are predicted to be in depression risk category j:

5. The depression risk screening optimization method based on large and small model linkage according to claim 1, characterized in that: Step S3 includes fine-tuning of the large language model, interaction design, and answer mapping mechanism.

6. The depression risk screening optimization method based on large and small model linkage according to claim 5, characterized in that: Fine-tuning and interaction design for large language models involves three phases: instruction fine-tuning, reinforcement learning optimization, and consistency constraint fine-tuning. These phases are performed sequentially, fine-tuning parameters at different levels. The instruction fine-tuning phase uses a fine-tuning dataset based on structured questions and answers generated by experts and templates to complete domain knowledge injection. This includes dataset acquisition, dynamic weight adjustment, and domain knowledge injection technologies. Dataset acquisition includes: collecting high-quality corpus in the field of mental health, annotating keyword matching, emotional intensity, question format, question path, etc. through experts to form a domain-specific fine-tuning dataset; manually integrating public literature databases to form an external medical database, and designing instruction templates for supervised training; Dynamic weight adjustment includes: on this basis, adjusting the loss function weights of different samples according to domain relevance and task complexity; Domain knowledge injection includes: using external medical knowledge bases (such as DSM-5 diagnostic criteria) to post-process and enhance the model; Specifically include: Design a weighted loss function L w , by assigning different weights w to different samples i To optimize model performance; the weighted loss function formula is as follows: Where N is the total number of samples, w i represents the weight of the i-th sample, which is determined by the depression risk correlation or depression risk level, L(y i ,y i ) is the loss function of a single sample; Weight w i The calculation can be based on the importance score s of the sample i : Among them, s i Calculated through psychiatric expert annotation and depression risk discrimination complexity indicators (such as question difficulty and user answer ambiguity); The reinforcement learning optimization phase uses real multi-round conversation data for fine-tuning training. The goal is to enable the model to generate responses that better meet user expectations, namely, the next question generation. This phase primarily includes: a context encoder, an intent recognition module, dialogue state tracking (DST), and a reinforcement learning optimization strategy. The context encoder uses the Transformer architecture built into the large language model to directly jointly model the user's historical conversations and current input; The intent recognition model uses pre-training to classify user input intent, supports fuzzy matching and multi-intent parsing, and understands the emotional tendencies (such as insomnia and loss of interest) in user responses. The conversation state tracking module is based on a memory network and dynamically updates the user's response status, including covered and uncovered emotional topics. The reinforcement learning optimization strategy is to introduce the reinforcement learning algorithm (PPO) to stop dialogue generation when the model completes the collection of standardized data features within a limited number of rounds; the goal of PPO is to minimize the agent objective function L CLIP : L CLIP (θ)=E t [min(r t (i)A t ,clip(r t (θ),1-ε,1+ε)A t )] (12) in, A represents the probability ratio of the trajectory of the new generated problem to the trajectory of the old problem, t is the advantage function, which indicates the advantage of the current question generation over the average normative question answering. ε is usually between 0.1 and 0.2, indicating the clipping range. clip(x,a,b) means clipping x to the interval [a,b]. Advantage function A t The formula is as follows: V(s t )A t =Q(s t ,a t )-V(s t ) (13) Among them, Q(s t ,a t ) represents the state-action value function, V(s t ) represents the state value function; Consistency constraint fine-tuning is the generation consistency control driven by the attention mechanism, including global attention constraints and local attention constraints; Specifically include: Global attention constraint: A global attention mechanism is introduced during the generation process to force the model to focus on the entire conversation history and questionnaire structure, ensuring that the generated content is consistent with the context. Consistency loss function: Design a consistency loss function based on contrastive learning to penalize the semantic deviation between the generated content and the reference answer, thereby improving the accuracy of the generated results; Among them, x i represents the anchor point sample, represents a positive sample (consistent with the anchor point), represents a negative sample (not consistent with the anchor point), d(x,y) represents the distance metric between samples (such as cosine similarity or Euclidean distance), y i represents the sample label (1 represents a positive sample, 0 represents a negative sample), and m is the boundary threshold; Local attention constraints: When generating each question, add local constraint modules (such as template matching and grammar checking) to ensure that the generated content conforms to language specifications and logical requirements; Autoregressive verification: After the content is generated and before it is delivered to the user, the generated content is retrospectively analyzed using an autoregressive verification mechanism by simulating multiple rounds of conversations. This detects potential logical conflicts or inconsistencies and automatically corrects them. Autoregressive prediction formula: Among them, x t Indicates that interaction with the user is the tth generated question, x ＜t represents the first t-1 generated questions; When checking for logical conflicts, you can calculate the KL divergence between the generated content and the reference question: Here, P(x) represents the distribution of generated content, and Q(x) represents the distribution of original reference questions.

7. The depression risk screening optimization method based on large and small model linkage according to claim 5, characterized in that: The answer mapping mechanism is based on a context encoder, which extracts the emotional intensity rating in the conversation, matches keywords, and aligns them with the questionnaire options, and builds an answer mapping rule library based on rules and statistical methods.

8. The depression risk screening optimization method based on large and small model linkage according to claim 1, characterized in that In step S4, after outputting the user's hierarchical outcomes and label annotations, the interactive interface sets the "approval" or "question" channel; When the user's feedback is "approval", the system marks it as "high-quality data" and stores it in the training database; When user feedback is "questionable," the "questionable" interaction content is marked as "data to be verified," triggering the model's self-correction mechanism to mine a better path from historical data. Detailed feedback (including specific questionable points and improvement suggestions) is recorded and stored in the training database after manual review or expert annotation, and the update mark is updated to "high-quality data." By analyzing feedback data, we extract key interaction indicators (such as emotional expression patterns, lifestyle descriptions, etc.) and incorporate them into feature engineering for the prediction model; Through regular statistical analysis, we can identify high-frequency problems or common demands, and guide the optimization of prediction logic and interaction methods.

9. The depression risk screening optimization method based on large and small model linkage according to claim 8, characterized in that: The aforementioned reinforcement learning of "high-quality data" is used for subsequent model fine-tuning, including: By giving positive reward signals, strengthening the current interaction strategy and prediction logic, and using high-quality data for incremental reinforcement learning training, the environment is defined: the state is the current conversation history, including all questions raised by the system and the user's answers, represented by a sequence: S t ={(q1,a1),(q2,a2),…,(q t ,a t )} (17) where q i is the i-th question, a i is the user's answer; action is the next question q generated by the system based on the current depression risk screening status t+1 ;Reward is the user's recognition of the final evaluation result; Use the optimization strategy network π θ (a t |s t ), that is, in a given state s t Next select action a t The probability distribution of maximizing the long-term cumulative reward; in each complete round of user research, the following conversation history is recorded: S t ={(q1,a1),(q2,a2),…,(q t ,a t )}; (18) Evaluation result y and user approval label R; reward function R(ST) is defined as: Use high-quality data to update the strategy. For each high-quality dialogue logic trajectory τ, calculate the gradient and update the parameters, where α is the learning rate:

10. A depression risk screening optimization system based on large and small model linkage, characterized in that: The system is used to execute the method according to any one of claims 1 to 9, and the system includes: The data collection end, which includes the scale source database, fine-tuning dataset, external medical database and training dataset, is used to collect and save data information and upload the data information to the cloud server; The cloud server is equipped with a data analysis module, a trained convolutional neural network model, a trained semantic analysis-enhanced fine-tuned large language model, and a trained human-computer interaction reinforcement learning model. The data analysis module includes algorithms for missing value interpolation, standardization, data format unification, basic statistical analysis, and feature engineering. The cloud server is used to pre-process the collected digital and text data, call the trained semantic analysis-enhanced fine-tuned large language model for dialogue generation and answer mapping, and further call the trained convolutional neural network model to predict the user's depression risk level. Feedback module, used to provide feedback to users on their evaluation of prediction results; A display module, used to display the depression risk level prediction results; The early warning module is used to remind users, provide explanatory explanations based on the depression risk level prediction results, and provide relief suggestions.

Images