Psychological Anxiety Risk Analysis Model Based on Large Language Model Interaction

Psychological Anxiety Risk Analysis Model Based on Large Language Model Interaction | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Psychological Anxiety Risk Analysis Model Based on Large Language Model Interaction Zhenyu Hou, Linjun Yan, Lailong Luo, Zhaoyun Ding, Yun Zhou This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8192788/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract In recent years, psychological anxiety has emerged as a pervasive mental health issue impacting socioeconomic development and individual well-being, making scientific assessment of anxiety cru- cial. While traditional manual evaluation methods and machine learning-based automated assessment approaches each possess unique advantages, they often struggle to simultaneously address efficiency, reliability, and feature generalization capabilities, lacking systematic comparative validation with standard clinical tools. To tackle this challenge, this study first proposes an anxiety risk assess- ment model based on interactive large language models (LLMs). The model is fine-tuned using multi-round psychological counseling dialogue datasets to provide high-quality psychological support. Subsequently, we introduce a dual-channel BERT framework for anxiety risk assessment, compris- ing a main channel and a high-risk channel. The framework strictly maps symptom dimensions and score ranges from the GAD-7 and HAMA clinical scales. This design enables accurate anxiety risk assessment while promoting psychological intervention through interactive dialogue. Experimental results demonstrate that the proposed model achieves high accuracy (88.86%) and robustness in gen- eral case evaluations, maintains 88.20% accuracy in cross-dataset validation, and exhibits superior performance in identifying high-risk cases (99.34% high-risk recall rate), significantly outperforming advanced baseline models like MentalBERT. psychological anxiety large language model (LLM) dynamic interaction risk assessment Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 1 Introduction Psychological anxiety has evolved into a widespread mental health concern worldwide, exerting substan- tial impacts on socioeconomic development and individual well-being. According to the World Health Organization (WHO) 2022 report, approximately 301 million people globally suffer from anxiety dis- orders. The Mental Health Blue Book: China’s National Psychological Health Development Report (2023–2024) indicates that the anxiety risk detection rate in China has reached 15.8%.Studies have shown that anxiety is closely associated with cognitive impairment [ 1 ] and may even lead to physiological problems such as cardiovascular diseases and immune dysfunction [ 2 ]. Traditional methods for assessing psychological anxiety mainly include self-rating scales (e.g., the GAD-7 scale) and structured clinical interviews. Self-rating scales possess high reliability and validity [ 3 ] as well as clinical convenience, but they are susceptible to social desirability bias, which may cause patients to deliberately downplay their symptoms [ 4 ]. Although clinical interviews are highly accurate, they are overly dependent on the experience of physicians, time-consuming, and difficult to cover large- scale populations [ 5 ]. In addition, traditional methods struggle to capture dynamic behavioral data and fail to reveal the potential associations between anxiety states, language expressions, and interaction behaviors. The origins of machine learning-based psychological anxiety risk analysis methods can be traced to the convergence of multiple disciplines, including psychology, computer science, and advancements in artificial intelligence. Early explorations began in the 1960s when computer scientist Joseph Weizenbaum developed the ELIZA chatbot [ 6 ], which pioneered attempts to simulate psychotherapeutic dialogues, laying the groundwork for subsequent affective computing and mental health analysis. In 1997, Pro- fessor Rosalind Picard at MIT formally introduced the concept of “Affective Computing”, marking the integration of artificial intelligence with emotion recognition technology and providing a theoreti- cal framework for automated psychological anxiety analysis. Since then, some scholars took the lead in introducing machine learning models such as support vector machines [ 7 ] and random forests into this field. By constructing multi-dimensional features, they achieved effective prediction of psychological risks like depression and confirmed the value of digital behavioral footprints in psychological assessment [ 8 ]. Meanwhile, there have also been studies that designed algorithms for different contexts [ 9 ] and multi- modal data to improve accuracy [ 10 ], yielding favorable results. Despite the certain achievements made in relevant research, there remains a limitation of insufficient feature generalization. With the proposal of the Transformer architecture, large language models (LLMs) have emerged. Compared with traditional language models, LLMs are of enormous scale, typically containing billions of parameters, and can understand and generate human-like conversational text with high precision [ 11 ], thus providing a new paradigm for mental health assessment. The application of large models in the field of mental health originated from unimodal text analysis, where the implementation of the BERT architecture enabled the identification of psychological states from text [ 12 ]. Subsequent studies have employed methods such as the CLIP architecture [ 13 ] and wav2vec 2.0 [ 14 ] to achieve multimodal anal- yses, such as video-text and speech-language analyses, further enhancing the accuracy of psychological state identification. Furthermore, numerous studies have demonstrated that LLMs exhibit great potential in early screening for depression, sentiment analysis, and the detection and classification of various men- tal health syndromes including social anxiety and suicidal ideation [ 15 ]. Meanwhile, when functioning as conversational agents (CAs) to provide digital mental health support [ 16 ], they possess advantages such as timeliness, non-judgmental attitude, and personalization, with some models even gaining recognition for their empathetic performance. However, this field still faces numerous challenges, such as performance degradation in non-English contexts due to scarce datasets [ 16 ], insufficient reliability of conversational agents in high-risk scenar- ios, limited memory affecting dialogue coherence, inconsistent outputs of LLMs influenced by prompt engineering, potential provision of inappropriate suggestions, and lack of necessary clinical judgment capabilities [ 17 ]. Therefore, this paper summarizes that existing studies mainly have the following three key limitations: 1) Over-reliance on static text data, neglecting temporal behavioral characteristics in interactions; 2) Poor model interpretability, restricting clinical application; 3) Lack of systematic solutions for data privacy and ethical risks. To address these challenges, this study proposes a LLM-based interactive model for psychological anxiety risk assessment. The model combines LLM fine-tuning with a dual-channel BERT-based anxiety risk evaluation system, enabling automated psychological interventions through conversational dialogue while assessing anxiety risks. Its evaluation criteria are rigorously developed using the GAD-7 and HAMA clinical scales to ensure alignment with clinical standards. The key contributions are as follows: • We design a hierarchical processing architecture (risk assessment layer and psychological support layer) to coordinate natural language dialogue and risk assessment. • We adopt LoRA low-rank adaptive fine-tuning with 4-bit quantization, achieving efficient inference (82 tokens/s) with only 3.125% parameter adjustments. • We develop a Dual-Channel BERT model that coordinates decision-making between the main classi- fication channel and high-risk channel, increasing the recall rate of high-risk samples by 5% compared with the baseline (F1-score = 0.8886). • We set up the mechanism of ethical intervention, designed the safety response strategy for high-risk scenarios, and defined the clinical auxiliary location of the model. 2 Related Works 2.1 Development and Limitations of Traditional Psychological Assessment Methods Research on psychological anxiety assessment began in the mid-20th century. In 1959, Hamilton [ 18 ] developed the Hamilton Anxiety Rating Scale (HAMA), establishing a standardized quantitative eval- uation system with an inter-rater reliability of 0.86, which has become the gold standard for clinical diagnosis. Subsequently, Zigmond and Snaith [ 19 ] designed the Hospital Anxiety and Depression Scale (HADS) for general hospitals, which eliminates interference from physical symptoms and achieves a Cron- bach’s α coefficient < 0.82 among patients with chronic diseases. In 2006, Spitzer et al. [ 3 ] developed the GAD-7 scale, which enables rapid screening through 7 concise questions with an AUC of 0.92. However, these methods have notable limitations. Lee et al. [ 20 ] found that HADS has a false negative rate of up to 34% in Asian populations. Brown et al. [ 21 ] pointed out that a full ADIS-5 interview takes 60–120 minutes, with a cost of $ 2,000 per person for clinicians, severely restricting its popularization. 2.2 Innovations and Breakthroughs of Machine Learning Methods In the early 21st century, computerized assessment became a research hotspot. In 2010, Kobak et al. [ 22 ] developed a Computer Adaptive Test (CAT) system based on Item Response Theory, which reduces test items by 40% while maintaining a reliability of α = 0.91. In 2012, Pestian et al. [ 23 ] pioneered the application of Support Vector Machines (SVM) in depression detection, achieving an F1-score of 0.68. In 2013, DeChoudhury et al. [ 24 ] improved the AUC to 0.79 by incorporating psycholinguistic features from the Linguistic Inquiry and Word Count (LIWC) tool.The academic community has also actively explored sentiment analysis methods tailored to the characteristics of different data types. In 2014, Li et al. [ 25 ] proposed a sentiment analysis algorithm suitable for short text data by combining the LDA model with the SVR model, achieving an optimal CC value of 0.64. Meanwhile, Ilikci et al. [ 26 ] employed heatmaps to train a YOLOv3 model for sentiment detection, demonstrating significantly faster processing speeds compared to ResNet and DenseNet. While machine learning has achieved remarkable success in psychological risk assessment, challenges such as feature redundancy continue to complicate computations, necessitating algorithm optimization for enhanced efficiency. Tadesse et al. [ 27 ] proposed a mutual information-based feature selection algorithm (MIFS) that reduces feature dimensions by 60% while maintaining model performance. Their study further revealed a significant positive correlation between first-person pronoun density (“I”) frequency and anxiety levels (r = 0.42, p < 0.01). Guntuku et al. [ 28 ] made algorithmic breakthroughs by integrating multi-source features through XGBoost + LightGBM, achieving an 83.5% accuracy rate in identifying high-risk individuals across 10,328 users from cross-platform studies. The model specifically optimized detection sensitivity for “smiling depression” (high-risk individuals with apparent positivity). 2.3 Paradigm Shifts in the Era of Large Language Models The emergence of the Transformer architecture [ 29 ] has brought fundamental changes. In 2019, Devlin et al. [ 12 ] applied BERT to mental health tasks, achieving an F1-score of 0.83 after fine-tuning on 120,000 clinical records. In 2021, Ji et al. [ 30 ] developed the domain-adaptive pre-trained model MentalBERT, Pearson correlation coefficient for GAD-7 prediction reached 0.89 (p < 0.001). Meanwhile, Large Language Models (LLMs) have also developed rapidly. A number of scholars have begun to attempt to apply LLMs to mental health tasks. In 2020, Fitzpatrick et al. [ 31 ] developed Woebot, which conducts interactive risk assessment through human-LLM dialogue and increases the emotional disclosure rate by 40% through dynamic dialogue. The study by Wang et al. [ 32 ] also proved that the empathetic responses generated by LLMs can significantly improve users’ emotional disclosure rate, thereby enhancing the recognition accuracy. Compared with traditional questionnaires, LLMs provide a more accurate, dynamic and flexible method for mental health risk assessment. However, challenges remain: Grabb [ 33 ] found 12% of LLM responses carry the risk of clinical mis- guidance in 2023; Vajre et al. [ 34 ] emphasized in 2023 that poor model interpretability limits clinical applicability. Current research shows three trends: 1) From static evaluation to dynamic interaction—Lee et al. [ 35 ] demonstrated in 2023 that empathetic responses increase emotional disclosure rates by 35%; 2) From single-modal to multimodal integration—Pampouchidou et al. [ 13 ] developed a video-text system in 2022 with a sensitivity of 91.3% for identifying latent symptoms; 3) From general-purpose models to domain-specific optimization, such as Huawei Cloud’s MindBot, which is dedicated to Chinese mental health services. These advances provide a foundation for our Dual-Channel evaluation model. 3 Methodology 3.1 Selection of the Base Model Currently, mainstream large language models mainly include ChatGPT, DeepSeek, Google Gemini, QwenLM, etc. Referring to the research by Rahman et al. [ 36 ] and combining the characteristics of anx- iety risk assessment tasks, this paper selects the following metrics to compare the performance of each model: MMLU (Overall): Using the Massive Multitask Language Understanding (MMLU) benchmark, it assesses factual recall, contextual reasoning, and complex question-answering capabilities across diverse knowledge domains. MMLU–Reasoning: The MMLU-Reasoning Subset evaluates a model’s ability to process multi-step, context-dependent questions within specialized domains like law or biology. HellaSwag: HellaSwag benchmarks commonsense reasoning by testing a model’s ability to predict the most plausible continuation of everyday scenarios. It measures intuitive understanding of physical and social contexts. MATH-500: The MATH-500 dataset assesses numerical and quantitative reasoning by requiring models to solve structured mathematical problems, including symbolic manipulation and equation solving. Scores reflect logical deduction efficiency. USMLE Step 1: The USMLE Step 1 benchmark evaluates medical knowledge integration, cover- ing pathophysiology, pharmacology, and diagnostics. Models are tested on clinical decision-making accuracy. CommonsenseQA: CommonsenseQA tests everyday reasoning through multiple-choice questions requiring implicit world knowledge. It highlights gaps in models’ real-world semantic understanding. The final results are shown in Figure.1. It can be found that for medical texts, the DeepSeek model has the best comprehensive performance with the least resource consumption. Therefore, this paper selects the DeepSeek-R1 model as the base model. 3.2 Model Training and Optimization We used DeepSeek-R1-Distill-Llama-8B as the base model and adopted 4-bit quantization (GPTQ algorithm) through the Unsloth optimization framework, achieving efficient operation with 4096 tokens maximum sequence length. The system implements dynamic conversation history management (maximum 8 dialogue rounds) with automatic context pruning to maintain optimal performance. Key hyperparameters were set as follows: Generation parameters: Primary response: temperature = 0.7; top-p = 0.9; repetition penalty = 1.1; Emotion analysis: temperature = 0.3 (deterministic sampling); Max new tokens: 800 for responses, 50 for emotion classification. Memory optimization: 4-bit quantization reducing memory demand to 15GB total utilization; Automatic history truncation preserving 16 most recent exchanges. Architecture features: Two-stage generation (emotion classification → contextual response); Explicit clinical reasoning via tags; Role-specific formatting (“client”/“Dr.Ivy”). The dialogue system demonstrates: real-time interaction speed maintained at < 2s latency per round; stable context utilization through 4096-token window management; clinical response quality via pre- response emotion analysis, structured reasoning prompts, and multi-round context preservation. Fine-tuning preserved the original 12.5% memory footprint of FP32 while extending capabilities for psychological counseling applications. The system shows particular effectiveness in maintaining natural dialogue flow despite reduced repetition penalty, delivering clinically-structured responses via template engineering, and balancing response diversity (top-p sampling) with consistency (temperature tuning). 3.3 Dual-Channel Risk Assessment Model We adopted an innovative Dual-Channel BERT architecture for anxiety risk assessment. A shared BERT- base feature extraction layer was used to obtain text semantic representations (hidden state H ∈ Rn ×768 ), with parallel operation of the main classification channel and high-risk detection channel.The main classification channel outputs probability distributions of low, moderate, and high anxiety through fully connected layers, as expressed in (1): Pmain = Softmax (Wm · Dropout (h[CLS]) + bm ) (1) The Dual-Channel BERT architecture we designed is shown in Fig. 2. The high-risk channel uses a deep compression structure (768→ 64→ 1) and a Sigmoid activation function to enhance the identification of high-risk features, as shown in (2): Phigh = σ (W2 · ReLU (W1 · h[CLS] )) (2) To address annotation noise, we designed a dynamic label smoothing strategy: 30% of high-risk boundary samples were re-annotated as moderate risk, and a weighting coefficient α = 0.7 was applied to the loss function, as shown in (3): where CE(·) denotes the cross-entropy loss, Ai is the annotation confidence, and yi is the true label.We used a Dual-Channel mechanism in the decision-making stage: If the highest class probability in the main channel was < 0.7 and the high-risk channel output exceeded 0.4, the sample was classified as high-risk; otherwise, the main channel result was retained. A weighted loss function (with λ = 2.0) was used to balance overall classification performance and high-risk recall, as shown in (4): Ltotal = Lmain + λ · BCE(Phigh, I[y = 2]) (4) where BCE(·) denotes the binary cross-entropy loss, and I[·] is an indicator function (1 if y = 2 (high risk), 0 otherwise). The model achieved a macro F1-score of 0.8886 on the test set, with an 3% increase in high-risk recall compared with the single-channel baseline. 3.4 System Integration and Real-Time Interaction We integrated the fine-tuned LLM with a Dual-Channel evaluation module to form an end-to-end interac- tive framework. During conversations, user inputs were encoded by BERT and processed simultaneously by the psychological support layer (generating CBT-based responses) and the risk assessment layer (cal- culating anxiety risks). Real-time risk assessment feedback was used to adjust the dialogue engine’s intervention strategies. The system had an average processing time of < 1 second per round and supported context tracking in over 15 rounds of extended dialogues. Its dynamic attention mechanism captured temporal behavioral patterns (e.g., emotional spikes or risk escalations). The output included natural language responses and visualized risk levels (low/moderate/high), with a clinical false positive rate < 5% (below the safety threshold). 4 Experiment 4.1 Datasets and Annotation Methods We used two key datasets for model training and validation: 1) The aiwei.json dataset: The aiwei.json dataset from the EmoLLM project (900 psychological coun- seling dialogues) was used for LLM fine-tuning. This dataset simulates professional counseling scenarios, with system prompts defining the model role as “Dr. Aiwei, a gentle senior nurse with rich psycho- logical expertise.” Each data entry includes user-reported psychological issues (input) and professional responses compliant with cognitive behavioral therapy (CBT) protocols (output), optimized using a thought chain template: “Step 1: Emotional recognition → Step 2: Empathetic expression → Step 3: Solution suggestions → Step 4: Safety verification.” 2) SoulChatCorpus-sft-multi-turn mixed dataset: This dataset is used to validate anxiety risk assess- ment. It contains 150,000 single-round consultations and 1 million multi-round conversations, with high-quality samples retained after rigorous privacy filtering. To address the lack of annotations, we developed a four-level classification system that strictly aligns with the clinical frameworks of the GAD-7 and HAMA scales, encompassing four dimensions: core symptoms, somatic symptoms, cognitive symp- toms, and functional impairment. The categories are defined as follows: No significant anxiety (using vague terms like ”somewhat,” corresponding to GAD-7 scores of 1–4); Mild anxiety (using absolute terms like ”always,” corresponding to GAD-7 scores of 5–9); Moderate anxiety (using explicit negative emotion expressions, corresponding to GAD-7 scores of 10–14); Severe anxiety (characterized by extreme expres- sions or severe somatic symptoms, corresponding to GAD-7 scores of 15–21). A team of 10 annotators underwent standardized training in GAD-7 criteria, conducted multiple rounds of annotation on the orig- inal samples, and implemented a dual-review mechanism. To verify annotation reliability, inter-annotator consistency was assessed using 500 randomly selected samples, yielding a Fleiss’ Kappa value of 0.82, indicating excellent annotation consistency. To address category imbalance in the original dataset, a core dataset was constructed based on clinical annotation standards. Targeted data augmentation strategies were then applied to generate final training/validation/testing data, including synonym substitution, random deletion (5%-15% non-keywords), and sentence restructuring. This produced 1,530 augmented samples for moderate anxiety and 1,707 for severe anxiety. The enhanced dataset ultimately contained 15,078 user messages with balanced category distribution: 5,041 for no significant anxiety, 6,023 for mild anxiety, 2,007 for moderate anxiety, and 2,007 for severe anxiety. A stratified random partitioning strat- egy was then employed to ensure balanced distribution across training, validation, and testing sets in a 14:3:3 ratio. 3)The generalization capability was validated using an external PsyDTCorpus dataset with identical annotation protocols, following the same labeling approach as the SoulChatCorpus-sft-multi-turn dataset. Inter-rater reliability was assessed across 500 randomly selected samples, yielding a Fleiss’ Kappa score of 0.85, demonstrating excellent annotation consistency. 4.2 Innovative Framework We proposed a “dynamic interaction + hierarchical evaluation” anxiety risk analysis paradigm, with the system architecture shown in Fig. 3 . Psychological Support Layer: A psychological counseling module using Chain-of-Thought (CoT) rea- soning decomposed traditional CBT into interpretable reasoning steps. It included an emotion state tracker to maintain the user’s real-time emotional vector space. Risk Assessment Layer: A Dual-Channel BERT method was used to improve the BERT BASE CHINESE model. It extracted anxiety semantics through a main classification channel, a high-risk clas- sification channel, and a fuzzy labeling method, with anxiety risk signals fed back to the psychological support layer. 4.3 Model Fine-tuning Performance We use Unsloth optimization framework to implement 4bit quantization training, and the key parameters of training are shown in Table 1 . Table 1 Training Parameters and Technical Advantages Parameter Value Technical Benefit Quantization Method GPTQ 4bit 62% VRAM reduction with maintained accuracy Batch Size 4 3.2 × throughput improvement Gradient Accumulation Steps 8 Enables large virtual batches Learning Rate 3 × 10 − 5 Optimized for 8B-scale convergence Optimizer AdamW-bnb-8bit Gradient quantization error < 0.8% The fine-tuned DeepSeek-R1-Distill-Llama-8B demonstrated superior efficiency, achieving 72 token- s/s inference speed with 15GB VRAM usage (4-bit quantized) while maintaining a 4096-token context window - outperforming comparable models like ChatGLM-6B (42 tokens/s, 13.2GB, 2048-token) and Qwen-1.8B (68 tokens/s, 5.8GB, 2048-token). As evidenced in Fig. 4 , our optimized training protocol maintained exceptional gradient stability, with norms converging to 0.51 ± 0.01 after 60 steps (vs. typ- ical 0.5-2.0 benchmarks), while Fig. 5 shows the loss descending from 2.1 to 1.1 (48% reduction) with logarithmic consistency (R2 = 0.97). This represents a 1.7 × speedup over baseline LLaMA2-7B (55 tokens/s, 18GB) despite our model’s larger context capacity, attributable to: 1) dynamic gradient scaling preventing norm fluctuations beyond the 0.5–0.75 range, and 2) Unsloth’s block-wise recomputation enabling stable 4-bit training. 4.4 Anxiety Risk Assessment Performance This study employs five leading models—BERT BASE CHINESE, MentalBERT, GLM-4-Air, CNN, and LSTM—as baselines. The proposed Dual-Channel BERT was tested against these models on the unified SoulChatCorpus-sft-multi-turn dataset for anxiety risk assessment. The classification performance differences are detailed in Table 2 and Table 3 . 4.4.1 Classification Result • BERT BASE CHINESE: The overall accuracy rate is 86.38%, and the weighted F1 score is also 86.38%. It performs relatively stable in the categories of no significant anxiety and mild anxiety. However, the Table 2 Classification Result Model Type True Label Predicted as none Predicted as low Predicted as mid Predicted as high BERT BASE CHINESE none 615 137 4 0 low 63 810 25 6 mid 0 48 242 11 high 0 2 12 287 Mental BERT none 650 99 5 2 low 84 780 35 5 mid 0 27 255 19 high 0 0 18 283 GLM-4 -Air none 583 169 4 0 low 48 574 277 5 mid 1 1 275 24 high 5 0 79 217 CNN none 559 195 1 1 low 73 811 18 2 mid 0 31 270 0 high 2 13 1 285 LSTM none 586 166 1 3 low 93 770 37 4 mid 0 31 268 2 high 0 5 3 293 Dual-Channel BERT none 640 112 3 1 low 87 796 20 1 mid 0 25 275 1 high 0 0 2 299 Table 3 Compare Experimental Results-SoulChatCorpus Model Type Precision Recall F1 Accuracy BERT BASE CHINESE 0.8673 0.8638 0.8638 0.8638 MentalBERT 0.8700 0.8703 0.8700 0.8700 GLM-4-Air 0.7892 0.7290 0.7419 0.7290 CNN 0.8589 0.8510 0.8509 0.8510 LSTM 0.8495 0.8475 0.8473 0.8475 Dual-Channel BERT 0.8886 0.8889 0.8886 0.8886 Table 4 Compare Experimental Results-PsyDTCorpus Model Type Precision Recall F1 Accuracy BERT BASE CHINESE 0.9175 0.9145 0.9153 0.9145 MentalBERT 0.9271 0.9273 0.9272 0.9273 GLM-4-Air 0.7882 0.7259 0.7392 0.7259 CNN 0.8961 0.8966 0.8962 0.8966 LSTM 0.8976 0.8972 0.8973 0.8972 Dual-Channel BERT 0.9319 0.9323 0.9320 0.9323 recall rate for moderate anxiety is only 80.40%, while the recall rate for severe anxiety reaches 95.35%, though there are still a few high-risk samples that were missed. MentalBERT: Its overall performance is comparable to BERT BASE CHINESE, achieving an accuracy rate of 87.00% and a weighted F1 score of 87.00%. It demonstrates good consistency in classifying no anxiety and mild anxiety, but falls slightly short of BERT BASE CHINESE in identifying moderate anxiety (81.47%) and severe anxiety (91.59%) with higher-risk categories. GLM-4-Air: The overall performance was the poorest,with an accuracy rate of only 72.90% and a weighted F1 score of 74.19%. There was significant category confusion, as mild anxiety was misclassified as moderate anxiety in 277 cases. The recall rate for severe anxiety was only 72.09%, failing to meet the reliability requirements for clinical screening. CNN achieved an overall accuracy of 85.10% and a weighted F1 score of 85.09%. The system demon- strated strong performance in classifying moderate and severe anxiety (with a F1 score of 96.77% for severe anxiety), but its recall rate for anxiety was only 73.94%, indicating a lack of distinct anxiety categories. Additionally, low-risk samples were frequently misclassified as mild anxiety. LSTM: Overall accuracy of 84.75% and weighted F1 score of 84.73%. It demonstrates outstanding per- formance in severe anxiety recognition (F1 score: 97.18%), but shows no significant confusion between anxiety and mild anxiety categories, with generally average classification stability. The Dual-Channel BERT model achieves optimal overall performance with an accuracy of 88.86% and a weighted F1 score of 88.86%. It demonstrates superior balance and reliability in anxiety category classification compared to the baseline model, particularly with a 99.34% recall rate for severe anxiety cases. The model exhibits an extremely low false negative rate for high-risk samples, fully meeting clinical intervention priorities. 4.4.2 Ablation Studies To evaluate the performance gains of the dual-channel architecture and fuzzy tag processing mechanism, an ablation experiment was conducted. Using the SoulChatCorpus-sft-multi-turn dataset, the study compared four models: the full Dual Channels BERT model, the model with removed fuzzy tags, the model with removed dual channels, and the model with both removed. The results are presented in Table 5 . When fuzzy labels are removed, the model’s precision, recall, F1 score, and accuracy all drop to 0.8703–0.8700. This demonstrates that the fuzzy label processing mechanism effectively mitigates category boundary ambiguity and enhances the model’s ability to capture semantic nuances. The absence of this mechanism would otherwise lead to reduced classification stability. Disabling dual channels (Remove Dual Channels) results in a performance drop (0.8783–0.8784), demonstrating that the dual-channel architecture optimizes feature extraction and classification logic through coordinated decision-making between the primary classification channel and high-risk channel. A single channel cannot adequately capture the complex semantic features of anxiety risk. Simultaneously removing both (Remove Dual Channels and Fuzzy Tags) reduced performance to 0.8673–0.8638, demonstrating the synergy between dual-channel and fuzzy-label mechanisms: fuzzy labels optimize annotation quality while dual channels enhance feature utilization, jointly ensuring the model’s accurate identification of varying anxiety risk levels. The full model (Dual Channels BERT) achieved precision, recall, F1 score, and accuracy of 0.8889–0.8886, significantly outperforming the post-ablation model. This demonstrates that the dual- channel architecture and fuzzy label processing mechanism are key design elements for enhancing anxiety risk assessment performance, effectively addressing core issues such as category confusion and high-risk misclassification. Table 5 Ablation Experiment Results Model Type Precision Recall F1 Accuracy Remove Fuzzy Tags 0.8703 0.8700 0.8700 0.8703 Remove Dual Channels 0.8783 0.8784 0.8783 0.8784 Remove Dual Channels and Fuzzy Tags 0.8673 0.8638 0.8638 0.8638 Dual Channels BERT 0.8889 0.8886 0.8886 0.8886 5 A Real-world Case Study To further validate our algorithm, we designed a user-friendly interactive interface that facilitates seam- less therapeutic interaction while providing real world anxiety monitoring. As illustrated in Fig. 6 , the interface adopts a dual-panel layout: the left panel hosts the conversational chatbot where users interact with virtual therapist ”Dr. Ava,” while the right panel displays analytical components including real- time anxiety scores, a dynamic trend chart tracking emotional fluctuations across conversation turns, and a comprehensive score distribution histogram. The interface incorporates a status indicator show- ing system connection status, along with a dedicated function for generating automated anxiety analysis reports summarizing key statistics (average score, maximum/minimum values, standard deviation) and trend interpretation. This design enables simultaneous therapeutic dialogue and quantitative mental state assessment, creating an immersive environment that balances clinical functionality with user-friendly interaction. To validate the practical effectiveness of our anxiety assessment system, we conducted a real-time case study using the AI Psychotherapy System interface Fig. 7 The system monitored a user initially expressing significant social anxiety regarding friendship loss and difficulties in making new connections. During the 8-round therapeutic dialogue, the user first revealed: ”I’m afraid of losing existing friend- ships and feel hesitant to actively make new friends.” Through structured guidance from virtual therapist ”Dr. Ava”, the conversation progressively addressed these concerns. The system’s dual-channel archi- tecture successfully detected subtle linguistic markers of anxiety while maintaining natural therapeutic interaction. As shown in the Anxiety Score Trend chart, the initial anxiety score reached 0.44 (round 2) when the user expressed core fears, but gradually decreased to 0.268 by round 8 after receiving cognitive restructuring and behavioral suggestions. The Score Distribution histogram demonstrates that 62.5% of conversation rounds maintained scores below 0.4, with only one round exceeding 0.44, indicating effective anxiety regulation during the ses- sion. This case vividly demonstrates our system’s comprehensive capability to first identify acute anxiety triggers in real-time through linguistic pattern recognition, then deliver appropriate therapeutic inter- ventions while continuously monitoring emotional fluctuations through its dual-channel architecture, and finally document the complete anxiety resolution journey from initial problem exposure to successful cog- nitive restructuring. The interface not only effectively supported the entire therapeutic process but also provided quantifiable anxiety metrics that tracked the user’s emotional progression, ultimately facilitat- ing a successful transition from avoidance behavior characterized by hesitation in making new friends to adopting proactive solutions such as joining tour groups for social connection. While the presented case demonstrates success, our system still has limitations including potential underestimation of intervention effects and risk assessment errors. When the system detects three con- secutive rounds of no significant reduction in user anxiety scores (fluctuation i0.05), it triggers a warning mechanism advising users: ”This issue may require professional face-to-face intervention. We recommend contacting local mental health services.” The system’s built-in error detection module automatically flags cases where high-risk assessments show below-threshold consistency with dialogue context (0.6), with reports noting ”This risk assessment requires further confirmation by professionals.” Example failure: A user describing ”recent frequent insomnia without noticeable impact on work or life” was misclassified as moderate anxiety (actual label: mild anxiety), reflecting the model’s excessive weighting of somatic symptoms. Future iterations will optimize the correlation modeling between symptoms and functional impacts. 6 Conclusion While traditional psychological anxiety risk assessment methods demonstrate good reliability and valid- ity, they require substantial clinical expertise from practitioners, limiting their scalability. These methods also face risks of patients deliberately concealing their conditions. Moreover, conventional approaches struggle to capture dynamic behavioral data or reveal potential correlations between anxiety states, ver- bal expressions, and interactive behaviors. This study develops a large-model-based psychological anxiety risk assessment system that evaluates users’ anxiety risks through monitoring their interactions with the model. The research focuses on three key aspects: Methodologically, we propose a ”dynamic interac- tion + hierarchical assessment” paradigm, with annotation standards strictly aligned to the GAD-7 and HAMA clinical scales, overcoming the static limitations of traditional assessments. Through LoRA fine- tuning and collaborative optimization of four-dimensional quantification, the system maintains clinical validity while reducing resource consumption by 87%, providing a solution for deployment in resource- constrained scenarios. Technically, the dual-channel BERT model significantly enhances high-risk sample identification through fuzzy label processing and collaborative decision-making mechanisms. Experiments show a 7.1% improvement in robustness under noisy labeling conditions and a 99.34% high-risk recall rate. In terms of application value, the end-to-end system achieves real-time assessment capability of processing 8.2 dialogue rounds per minute, supporting dynamic tracking of anxiety risks. By designing ethical intervention mechanisms and clarifying clinical assistance positioning, the system reduces misuse risks in high-risk scenarios. By encoding clinical criteria into an interpretable dialogue process (PRDS strategy), this model ensures interventions comply with CBT treatment protocols. Experimental results demonstrate that the proposed model exhibits high accuracy and robustness in anxiety risk assessment, effectively complementing traditional scales and clinical interviews while providing new technical sup- port for digital mental health interventions. The system is designed for risk assessment and early support rather than crisis intervention. When detecting high-risk expressions like suicidal ideation, it offers empa- thetic support while strongly recommending professional help and stopping further discussion to prevent potential harm. System performance is limited by training data and lacks actual emergency response capabilities. All assessment results should be used as clinical references and cannot replace professional diagnosis. 6.1 Future Work Despite the encouraging outcomes of this study in psychological anxiety risk identification, several limi- tations should be acknowledged, pointing toward meaningful directions for future research. One primary area for improvement lies in the enhancement and diversification of training data. The current dataset may not adequately represent all demographic groups or mental health conditions, which constrains the model’s interpretability and robustness. Subsequent efforts would benefit from incorporating more diverse data sources—such as social media content, news articles, and medical platform records—to improve generalizability and practical applicability. Furthermore, the reliance solely on textual data limits the model’s ability to capture non-verbal behavioral cues. Developing a multimodal evaluation framework that integrates audio, visual, and other forms of data could significantly enhance detection accuracy by offering a more comprehensive view of user expressions. Next, we will conduct prospective clinical validation studies through parallel assessments with psy- chiatrists to further validate the diagnostic accuracy, reliability, and practicality of the clinical utility verification model in real-world clinical settings. This will clarify its applicability as an auxiliary screening tool, establish operational protocols and quality control standards, and promote the seamless integration of digital assessment tools with traditional psychiatric diagnostic workflows. Another promising direction involves extending the model’s capability to multiple languages and cultural contexts. While the current system performs effectively in Chinese settings, its performance degrades when applied to other languages. Given the global and culturally varied nature of mental health needs, future work should prioritize training on multilingual and culturally diverse datasets. Such efforts would foster greater inclusivity and improve the model’s sensitivity to linguistic and cultural nuances, enabling more reliable anxiety risk detection across different populations. In conclusion, the anxiety risk identification model developed in this study provides a useful supple- mentary tool for round-the-clock mental health monitoring through conversational interactions. It holds promise as a scalable aid to existing clinical screening practices. Nonetheless, continued innovation and refinement are essential to broaden its applicability and strengthen its real-world impact. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Availability of data and material The datasets analysed during the current study are available in the EmoLLM (aiwei.json) and SoulChatCorpus repositories on GitHub, accessible at: • https://github.com/SmartFlowAI/EmoLLM/blob/main/datasets/aiwei.json • https://github.com/scutcyr/SoulChat Competing interests The authors declare that they have no competing interests. Funding Not applicable. Authors’ contributions Y. L. J. and H. Z. Y. were responsible for the writing of the original draft; L. L. L., D. Z. Y. and Z. Y. critically reviewed and edited the manuscript; Z. Y. served as the corresponding author, handling all correspondence and coordination with the journal during submission and peer review. All authors have read and approved the final submitted version and agree to be accountable for all aspects of the work. Acknowledgements Not applicable. References Hirsch CR, Mathews A. A cognitive model of pathological worry. Behav Res Ther. 2012;50(10):636–46. Chand SP, Marwaha R, Bender RM. Anxiety (nursing) (2021). Spitzer RL, Kroenke K, Williams JB, Lo… we B. A brief measure for assessing generalized anxiety disorder: the gad-7. Arch Intern Med. 2006;166(10):1092–7. Plummer F, Manea L, Trepel D, McMillan D. Screening for anxiety disorders with the gad-7 and gad-2: a systematic review and diagnostic metaanalysis. Gen Hosp Psychiatry. 2016;39:24–31. Torous J, Myrick KJ, Rauseo-Ricupero N, Firth J, et al. Digital mental health and covid-19: using technology today to accelerate the curve on access and quality tomorrow. JMIR mental health. 2020;7(3):18848. Weizenbaum J. Eliza—a computer program for the study of natural language communication between man and machine. Commun ACM. 1966;9(1):36–45. Orru G, Pettersson-Yeo W, Marquand AF, Sartori G, Mechelli A. Using support vector machine to identify imaging biomarkers of neurological and psychiatric disease: a critical review. Neurosci Biobehavioral Reviews. 2012;36(4):1140–52. De Choudhury M, Gamon M, Counts S, Horvitz E. Predicting depression via social media. In: Proceedings of the International AAAI Conference on Web and Social Media, vol. 7, pp. 128–137 (2013). Zhao W, Cai Z, Dong S, Hei W. A review of machine learning-based assessment of depression. In: China Health Information Processing Conference, pp. 266–290 (2024). Springer. Khoo LS, Lim MK, Chong CY, McNaney R. Machine learning for multimodal mental health detection: a systematic review of passive sensing approaches. Sensors. 2024;24(2):348. Naveed H, Khan AU, Qiu S, Saqib M, Anwar S, Usman M, Akhtar N, Barnes N, Mian A. A comprehensive overview of large language models. ACM Transactions on Intelligent Systems and Technology; 2023. Devlin J, Chang M-W, Lee K, Toutanova K: Bert: Pre-training of deep bidirectional trans- formers for language understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (long and, Papers S.), pp. 4171–4186 (2019). Pampouchidou A, Pediaditis M, Kazantzaki E, Sfakianakis S, Apostolaki I-A, Argyraki K, Manousos D, Meriaudeau F, Marias K, Yang F, et al. Automated facial video-based recog- nition of depression and anxiety symptom severity: cross-corpus validation. Mach Vis Appl. 2020;31(4):30. Diep B, Stanojevic M, Novikova J. Multi-modal deep learning system for depression and anxiety detection. arXiv preprint arXiv:221214490 (2022). Danner M, Hadzic B, Gerhardt S, Ludwig S, Uslu I, Shao P, Weber T, Shiban Y, Ratsch M. Advancing mental health diagnostics: Gpt-based method for depression detection. In: 2023 62nd Annual Conference of the Society of Instrument and Control Engineers (SICE), pp. 1290–1296 (2023). IEEE. Guo Z, Lai A, Thygesen JH, Farrington J, Keen T, Li K, et al. Large language models for mental health applications: systematic review. JMIR mental health. 2024;11(1):57400. Perlis RH, Goldberg JF, Ostacher MJ, Schneck CD. Clinical decision support for bipolar depression using large language models. Neuropsychopharmacology. 2024;49(9):1412–6. Hamilton M. The assessment of anxiety states by rating. Br J Med Psychol (1959). Zigmond AS, Snaith RP. The hospital anxiety and depression scale. Acta psychiatrica scandi- navica. 1983;67(6):361–70. Stafford L, Berk M, Jackson HJ. Validity of the hospital anxiety and depression scale and patient health questionnaire-9 to screen for depression in patients with coronary artery disease. Gen Hosp Psychiatry. 2007;29(5):417–24. Hasin DS, Greenstein E, Aivadyan C, Stohl M, Aharonovich E, Saha T, Goldstein R, Nunes EV, Jung J, Zhang H, et al. The alcohol use disorder and associated disabilities interview schedule-5 (audadis-5): procedural validity of substance use disorders modules through clinical re- appraisal in a general population sample. Drug Alcohol Depend. 2015;148:40–6. Kaufman J, Kobak K, Birmaher B, Lacy N. Ksads-comp perspectives on child psychiatric diagnostic assessment and treatment planning. J Am Acad Child Adolesc Psychiatry. 2020;60(5):540–2. Pestian JP, Matykiewicz P, Linn-Gust M, South B, Uzuner O, Wiebe J, Cohen KB, Hurdle J, Brew C. Sentiment analysis of suicide notes: A shared task. Biomedical Inf insights. 2012;5:9042. De Choudhury M, Counts S, Horvitz EJ, Hoff A. Characterizing and predicting postpartum depression from shared facebook data. In: Proceedings of the 17th ACM Conference on Computer Supported Cooperative Work & Social Computing, pp. 626–638 (2014). Li D, Niu J, Qiu M, Liu M. Sentiment analysis on weibo data. In: 2014 IEEE Computers, Communications and IT Applications Conference, pp. 249–254 (2014). IEEE. Ilikci B, Chen L, Cho H, Liu Q. Heat-map based emotion and face recognition from thermal images. In: 2019 Computing, Communications and IoT Applications (ComComAp), pp. 449–453 (2019). IEEE. Tadesse MM, Lin H, Xu B, Yang L. Detection of depression-related posts in reddit social media forum. IEEE Access. 2019;7:44883–93. Guntuku SC, Yaden DB, Kern ML, Ungar LH, Eichstaedt JC. Detecting depression and mental illness on social media: an integrative review. Curr Opin Behav Sci. 2017;18:43–9. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser L–. Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017). Ji S, Zhang T, Ansari L, Fu J, Tiwari P, Cambria E. Mentalbert: Publicly available pretrained language models for mental healthcare. arXiv preprint arXiv:211015621 (2021). Fitzpatrick KK, Darcy A, Vierhile M. Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (woebot): a randomized controlled trial. JMIR mental health. 2017;4(2):7785. Wang X, Sharma D, Kumar D. A review on ai-based modeling of empathetic conversational response generation. In: 2023 Asia Conference on Cognitive Engineering and Intelligent Interaction (CEII), pp. 102–109 (2023). IEEE. Grabb D. The impact of prompt engineering in large language model performance: a psychiatric example. J Med Artif Intell 6 (2023). Zhang W, Deng Y, Liu B, Pan SJ, Bing L. Sentiment analysis in the era of large language models: A reality check. arXiv preprint arXiv:230515005 (2023). Ding Y, Liu J, Zhang X, Yang Z. Dynamic tracking of state anxiety via multi-modal data and machine learning. Front Psychiatry. 2022;13:757961. Rahman A, Mahir SH, Tashrif MTA, Aishi AA, Karim MA, Kundu D, Debnath T, Moududi MAA, Eidmum M. Comparative analysis based on deepseek, chatgpt, and google gemini: Features, techniques, performance, future prospects. arXiv preprint arXiv:250304783 (2025). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8192788","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":554364249,"identity":"0a50d3d6-e2d3-477d-93ef-354710cd7e6d","order_by":0,"name":"Zhenyu Hou","email":"","orcid":"","institution":"National University of Defense Technology","correspondingAuthor":false,"prefix":"","firstName":"Zhenyu","middleName":"","lastName":"Hou","suffix":""},{"id":554364252,"identity":"fef5bb57-0f05-47b1-bc5f-7acb9fbff4ac","order_by":1,"name":"Linjun Yan","email":"","orcid":"","institution":"National University of Defense Technology","correspondingAuthor":false,"prefix":"","firstName":"Linjun","middleName":"","lastName":"Yan","suffix":""},{"id":554364254,"identity":"6361f76c-8492-4dc9-9593-6500a51c5b0c","order_by":2,"name":"Lailong Luo","email":"","orcid":"","institution":"National University of Defense Technology","correspondingAuthor":false,"prefix":"","firstName":"Lailong","middleName":"","lastName":"Luo","suffix":""},{"id":554364255,"identity":"efa1689a-9166-48d2-adbd-f27422eb2342","order_by":3,"name":"Zhaoyun Ding","email":"","orcid":"","institution":"National University of Defense Technology","correspondingAuthor":false,"prefix":"","firstName":"Zhaoyun","middleName":"","lastName":"Ding","suffix":""},{"id":554364256,"identity":"674a373f-6f76-4dad-8c0e-8d8a0d9fa09a","order_by":4,"name":"Yun Zhou","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzUlEQVRIiWNgGAWjYDACCQglB8QGQMxMvBZjIDZsIElLYgPRWuRnNx+T5m27k75dInn7A4YK68QG9rMH8GphnHMsDajlWe7OGWmFDQxn0hMbePIS8GphlsgxA2o5nLvhRo5hA2Pb4cQGCR4DvFrYoFrSDcBa/hGhhQeqJQGipYEILRISacmWc84dNtxw5lnhjIRj6cZtPDn4tcjPSD54403ZYXmD48kbPnyosZbtZz+DXwsIMPHAWAkg3xFUDwSMP4hRNQpGwSgYBSMXAAD/UUOgBNp/eAAAAABJRU5ErkJggg==","orcid":"","institution":"National University of Defense Technology","correspondingAuthor":true,"prefix":"","firstName":"Yun","middleName":"","lastName":"Zhou","suffix":""}],"badges":[],"createdAt":"2025-11-24 11:23:21","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8192788/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8192788/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":97489062,"identity":"6ceffc89-fba1-4641-bb30-ae727e4e504b","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":373293,"visible":true,"origin":"","legend":"","description":"","filename":"PsychologicalAnxietyRiskAnalysisModelbasedonLLMs.docx","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/9e584a041788db05e845f4f9.docx"},{"id":97670849,"identity":"1b8c15a4-9015-4a9e-b96b-4fc8569d885a","added_by":"auto","created_at":"2025-12-08 09:31:26","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6941,"visible":true,"origin":"","legend":"","description":"","filename":"0429425bc92f483c91cb3499decf94c1.json","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/c1031bcdc7b7b1d1fa850af8.json"},{"id":97489047,"identity":"981b0ce0-614a-4ef6-ac0e-3ded1b715d19","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":125859,"visible":true,"origin":"","legend":"","description":"","filename":"0429425bc92f483c91cb3499decf94c11enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/f21cab45fa60faddb7701f39.xml"},{"id":97670467,"identity":"22f424a3-9fad-46fa-99b9-f4f57ad387b8","added_by":"auto","created_at":"2025-12-08 09:30:45","extension":"jpeg","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":300848,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/abff3ed3d74e8d3a30d40146.jpeg"},{"id":97671041,"identity":"01a6e884-4fc4-4c92-835a-3e642db891de","added_by":"auto","created_at":"2025-12-08 09:31:47","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6236,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/5d92e3941a279737681e354f.png"},{"id":97489055,"identity":"fab0c433-7d05-47d6-9972-9587e80946c6","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":37637,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/98785710e3e91ee2723e8b1b.png"},{"id":97669108,"identity":"88188452-7cac-497d-86de-311cabd3fdf6","added_by":"auto","created_at":"2025-12-08 09:27:14","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":71185,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/52804c39bc28f0720f306132.jpeg"},{"id":97489054,"identity":"c5596adb-8e17-40ab-8c8a-225ce6f8a9c1","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13220,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/764f209dfaa80519d54ddfb3.png"},{"id":97669172,"identity":"1fa9d4d4-05b0-4edc-be6a-fc1e5897c1b3","added_by":"auto","created_at":"2025-12-08 09:27:31","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15602,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/434059b62a3ac9b8333afcbf.png"},{"id":97489060,"identity":"9d068749-d921-4dd0-a2ea-b26ab1f1fe5d","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":43819,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/6e7417b2a19c44e8b44e5715.png"},{"id":97669974,"identity":"9999970e-608f-4268-a024-63a96915f6c7","added_by":"auto","created_at":"2025-12-08 09:29:28","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":78249,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/ef42234c16eb2dc696e11ebb.png"},{"id":97669472,"identity":"46ca632a-5473-45df-822a-5be0333cbff8","added_by":"auto","created_at":"2025-12-08 09:28:04","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":83163,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/f4c2b117f180c52bded8cec3.png"},{"id":97489058,"identity":"7ee0ff89-8468-4803-983d-6f230d100eaf","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1322,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/22b070849c67fb09cfd92bff.png"},{"id":97670389,"identity":"4518232b-4899-4001-9559-bc2aa92538d6","added_by":"auto","created_at":"2025-12-08 09:30:28","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9580,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/9a4311c652fcb182b87e7f11.png"},{"id":97670997,"identity":"6d52616c-1f21-4864-a642-ccd270b9a8c2","added_by":"auto","created_at":"2025-12-08 09:31:42","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":26479,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/77b2d568b251587920035172.png"},{"id":97489059,"identity":"6c31000a-072a-46e1-a655-180fcf3b748d","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3084,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/6708682dafeff5cfb98b307d.png"},{"id":97489065,"identity":"786ee587-dd6a-400f-a02e-edf1a9c52269","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3474,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/872652381a341c7a8515d6bb.png"},{"id":97489070,"identity":"1cc4e6cb-5ee8-4fe8-ad75-d530a1ef364b","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4819,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/aa7f9e54eb9461829d057f76.png"},{"id":97670494,"identity":"c1c7b1eb-e12e-4848-b367-5aeb0f721adc","added_by":"auto","created_at":"2025-12-08 09:30:48","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9174,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/41f3e45704f427e808e2e5f7.png"},{"id":97669155,"identity":"c4608ab6-c81a-49ef-8d65-7a603baef0bd","added_by":"auto","created_at":"2025-12-08 09:27:26","extension":"xml","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":123391,"visible":true,"origin":"","legend":"","description":"","filename":"0429425bc92f483c91cb3499decf94c11structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/b35670fac800ea977c4526cd.xml"},{"id":97489072,"identity":"85bf7c2e-bff6-41c9-b3ea-2aa491c3e1ea","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"html","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":132709,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/088be611fb980baf25714d25.html"},{"id":97669722,"identity":"9aad8593-4287-4c13-b64d-11123c8f30ef","added_by":"auto","created_at":"2025-12-08 09:28:46","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":218233,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of the performance of various models\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/030e195946fdb32b08ffeb9e.png"},{"id":97670607,"identity":"e67cd9c0-6242-4d2d-9440-61edcd804264","added_by":"auto","created_at":"2025-12-08 09:31:01","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":131282,"visible":true,"origin":"","legend":"\u003cp\u003eDual-Channel BERT architecture\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/c5b8336d6f99d6fa6eb07d36.png"},{"id":97489044,"identity":"a7e9a7ed-c909-4342-b0d9-22a11c851623","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":51935,"visible":true,"origin":"","legend":"\u003cp\u003eHierarchical evaluation architecture diagram.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/bd5b0fbf62ff473c277d68ee.png"},{"id":97489052,"identity":"2f752a5a-8810-401c-81f3-a8fffa768631","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":35352,"visible":true,"origin":"","legend":"\u003cp\u003eThe model loss function changes with model training iterations from 20 to 100\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/99356d5edbf2bbcaa1d3ca64.png"},{"id":97489048,"identity":"729d967b-67c4-4825-bd1b-11da2a41ec6a","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":37172,"visible":true,"origin":"","legend":"\u003cp\u003eGradient norm variation with model training iterations from 20 to 100\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/23c29c276833823fa81bbb17.png"},{"id":97489050,"identity":"937a1d4b-36d0-4e06-9641-99c8ace39db8","added_by":"auto","created_at":"2025-12-05 01:57:07","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":84846,"visible":true,"origin":"","legend":"\u003cp\u003eDeployed web interface\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/7d62230b308fd4738a881f30.png"},{"id":97669150,"identity":"5a85df09-3fdc-48f6-b7c0-774dc110239d","added_by":"auto","created_at":"2025-12-08 09:27:26","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":114381,"visible":true,"origin":"","legend":"\u003cp\u003eValidate system effectiveness through case studies\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/53b45a93ea694e18c0a4c3f0.png"},{"id":97677819,"identity":"4387c4ba-5525-42dd-b1ec-8436995e9818","added_by":"auto","created_at":"2025-12-08 09:54:41","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1617586,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8192788/v1/bc2aadd4-7d6c-4fd7-abc8-bc372aa75d5f.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Psychological Anxiety Risk Analysis Model Based on Large Language Model Interaction","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003ePsychological anxiety has evolved into a widespread mental health concern worldwide, exerting substan- tial impacts on socioeconomic development and individual well-being. According to the World Health Organization (WHO) 2022 report, approximately 301\u0026nbsp;million people globally suffer from anxiety dis- orders. The Mental Health Blue Book: China\u0026rsquo;s National Psychological Health Development Report (2023\u0026ndash;2024) indicates that the anxiety risk detection rate in China has reached 15.8%.Studies have shown that anxiety is closely associated with cognitive impairment [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] and may even lead to physiological problems such as cardiovascular diseases and immune dysfunction [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTraditional methods for assessing psychological anxiety mainly include self-rating scales (e.g., the GAD-7 scale) and structured clinical interviews. Self-rating scales possess high reliability and validity [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] as well as clinical convenience, but they are susceptible to social desirability bias, which may cause patients to deliberately downplay their symptoms [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Although clinical interviews are highly accurate, they are overly dependent on the experience of physicians, time-consuming, and difficult to cover large- scale populations [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. In addition, traditional methods struggle to capture dynamic behavioral data and\u003c/p\u003e\u003cp\u003efail to reveal the potential associations between anxiety states, language expressions, and interaction behaviors.\u003c/p\u003e\u003cp\u003eThe origins of machine learning-based psychological anxiety risk analysis methods can be traced to the convergence of multiple disciplines, including psychology, computer science, and advancements in artificial intelligence. Early explorations began in the 1960s when computer scientist Joseph Weizenbaum developed the ELIZA chatbot [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], which pioneered attempts to simulate psychotherapeutic dialogues, laying the groundwork for subsequent affective computing and mental health analysis. In 1997, Pro- fessor Rosalind Picard at MIT formally introduced the concept of \u0026ldquo;Affective Computing\u0026rdquo;, marking the integration of artificial intelligence with emotion recognition technology and providing a theoreti- cal framework for automated psychological anxiety analysis. Since then, some scholars took the lead in introducing machine learning models such as support vector machines [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] and random forests into this field. By constructing multi-dimensional features, they achieved effective prediction of psychological risks like depression and confirmed the value of digital behavioral footprints in psychological assessment [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Meanwhile, there have also been studies that designed algorithms for different contexts [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] and multi- modal data to improve accuracy [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], yielding favorable results. Despite the certain achievements made in relevant research, there remains a limitation of insufficient feature generalization.\u003c/p\u003e\u003cp\u003eWith the proposal of the Transformer architecture, large language models (LLMs) have emerged. Compared with traditional language models, LLMs are of enormous scale, typically containing billions of parameters, and can understand and generate human-like conversational text with high precision [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], thus providing a new paradigm for mental health assessment. The application of large models in the field of mental health originated from unimodal text analysis, where the implementation of the BERT architecture enabled the identification of psychological states from text [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Subsequent studies have employed methods such as the CLIP architecture [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] and wav2vec 2.0 [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] to achieve multimodal anal- yses, such as video-text and speech-language analyses, further enhancing the accuracy of psychological state identification. Furthermore, numerous studies have demonstrated that LLMs exhibit great potential in early screening for depression, sentiment analysis, and the detection and classification of various men- tal health syndromes including social anxiety and suicidal ideation [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Meanwhile, when functioning as conversational agents (CAs) to provide digital mental health support [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], they possess advantages such as timeliness, non-judgmental attitude, and personalization, with some models even gaining recognition for their empathetic performance.\u003c/p\u003e\u003cp\u003eHowever, this field still faces numerous challenges, such as performance degradation in non-English contexts due to scarce datasets [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], insufficient reliability of conversational agents in high-risk scenar- ios, limited memory affecting dialogue coherence, inconsistent outputs of LLMs influenced by prompt engineering, potential provision of inappropriate suggestions, and lack of necessary clinical judgment capabilities [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Therefore, this paper summarizes that existing studies mainly have the following three key limitations: 1) Over-reliance on static text data, neglecting temporal behavioral characteristics in interactions; 2) Poor model interpretability, restricting clinical application; 3) Lack of systematic solutions for data privacy and ethical risks.\u003c/p\u003e\u003cp\u003eTo address these challenges, this study proposes a LLM-based interactive model for psychological anxiety risk assessment. The model combines LLM fine-tuning with a dual-channel BERT-based anxiety risk evaluation system, enabling automated psychological interventions through conversational dialogue while assessing anxiety risks. Its evaluation criteria are rigorously developed using the GAD-7 and HAMA clinical scales to ensure alignment with clinical standards. The key contributions are as follows:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u0026bull; We design a hierarchical processing architecture (risk assessment layer and psychological support layer) to coordinate natural language dialogue and risk assessment.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u0026bull; We adopt LoRA low-rank adaptive fine-tuning with 4-bit quantization, achieving efficient inference (82 tokens/s) with only 3.125% parameter adjustments.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u0026bull; We develop a Dual-Channel BERT model that coordinates decision-making between the main classi- fication channel and high-risk channel, increasing the recall rate of high-risk samples by 5% compared with the baseline (F1-score\u0026thinsp;=\u0026thinsp;0.8886).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e \u0026bull; We set up the mechanism of ethical intervention, designed the safety response strategy for high-risk scenarios, and defined the clinical auxiliary location of the model.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e"},{"header":"2 Related Works","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Development and Limitations of Traditional Psychological Assessment Methods\u003c/h2\u003e\u003cp\u003eResearch on psychological anxiety assessment began in the mid-20th century. In 1959, Hamilton [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] developed the Hamilton Anxiety Rating Scale (HAMA), establishing a standardized quantitative eval- uation system with an inter-rater reliability of 0.86, which has become the gold standard for clinical diagnosis. Subsequently, Zigmond and Snaith [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] designed the Hospital Anxiety and Depression Scale (HADS) for general hospitals, which eliminates interference from physical symptoms and achieves a Cron- bach\u0026rsquo;s α coefficient\u0026thinsp;\u0026lt;\u0026thinsp;0.82 among patients with chronic diseases. In 2006, Spitzer et al. [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] developed the GAD-7 scale, which enables rapid screening through 7 concise questions with an AUC of 0.92.\u003c/p\u003e\u003cp\u003eHowever, these methods have notable limitations. Lee et al. [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] found that HADS has a false negative rate of up to 34% in Asian populations. Brown et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] pointed out that a full ADIS-5 interview takes 60\u0026ndash;120 minutes, with a cost of \u003cspan\u003e$\u003c/span\u003e2,000 per person for clinicians, severely restricting its popularization.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Innovations and Breakthroughs of Machine Learning Methods\u003c/h2\u003e\u003cp\u003eIn the early 21st century, computerized assessment became a research hotspot. In 2010, Kobak et al.\u003c/p\u003e\u003cp\u003e[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] developed a Computer Adaptive Test (CAT) system based on Item Response Theory, which reduces test items by 40% while maintaining a reliability of α\u0026thinsp;=\u0026thinsp;0.91. In 2012, Pestian et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] pioneered the application of Support Vector Machines (SVM) in depression detection, achieving an F1-score of 0.68. In 2013, DeChoudhury et al. [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] improved the AUC to 0.79 by incorporating psycholinguistic features from the Linguistic Inquiry and Word Count (LIWC) tool.The academic community has also actively explored sentiment analysis methods tailored to the characteristics of different data types. In 2014, Li et al. [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] proposed a sentiment analysis algorithm suitable for short text data by combining the LDA model with the SVR model, achieving an optimal CC value of 0.64. Meanwhile, Ilikci et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] employed heatmaps to train a YOLOv3 model for sentiment detection, demonstrating significantly faster processing speeds compared to ResNet and DenseNet.\u003c/p\u003e\u003cp\u003eWhile machine learning has achieved remarkable success in psychological risk assessment, challenges such as feature redundancy continue to complicate computations, necessitating algorithm optimization for enhanced efficiency. Tadesse et al. [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] proposed a mutual information-based feature selection algorithm (MIFS) that reduces feature dimensions by 60% while maintaining model performance. Their study further revealed a significant positive correlation between first-person pronoun density (\u0026ldquo;I\u0026rdquo;) frequency and anxiety levels (r\u0026thinsp;=\u0026thinsp;0.42, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). Guntuku et al. [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] made algorithmic breakthroughs by integrating multi-source features through XGBoost\u0026thinsp;+\u0026thinsp;LightGBM, achieving an 83.5% accuracy rate in identifying high-risk individuals across 10,328 users from cross-platform studies. The model specifically optimized detection sensitivity for \u0026ldquo;smiling depression\u0026rdquo; (high-risk individuals with apparent positivity).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Paradigm Shifts in the Era of Large Language Models\u003c/h2\u003e\u003cp\u003eThe emergence of the Transformer architecture [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] has brought fundamental changes. In 2019, Devlin et al. [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] applied BERT to mental health tasks, achieving an F1-score of 0.83 after fine-tuning on 120,000 clinical records. In 2021, Ji et al. [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e] developed the domain-adaptive pre-trained model MentalBERT, Pearson correlation coefficient for GAD-7 prediction reached 0.89 (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Meanwhile, Large Language Models (LLMs) have also developed rapidly. A number of scholars have begun to attempt to apply LLMs to mental health tasks. In 2020, Fitzpatrick et al. [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] developed Woebot, which conducts interactive risk assessment through human-LLM dialogue and increases the emotional disclosure rate by 40% through dynamic dialogue. The study by Wang et al. [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e] also proved that the empathetic responses generated by LLMs can significantly improve users\u0026rsquo; emotional disclosure rate, thereby enhancing the recognition accuracy. Compared with traditional questionnaires, LLMs provide a more accurate, dynamic and flexible method for mental health risk assessment.\u003c/p\u003e\u003cp\u003eHowever, challenges remain: Grabb [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] found 12% of LLM responses carry the risk of clinical mis- guidance in 2023; Vajre et al. [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] emphasized in 2023 that poor model interpretability limits clinical applicability.\u003c/p\u003e\u003cp\u003eCurrent research shows three trends: 1) From static evaluation to dynamic interaction\u0026mdash;Lee et al.\u003c/p\u003e\u003cp\u003e[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e] demonstrated in 2023 that empathetic responses increase emotional disclosure rates by 35%; 2) From single-modal to multimodal integration\u0026mdash;Pampouchidou et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] developed a video-text system in 2022 with a sensitivity of 91.3% for identifying latent symptoms; 3) From general-purpose models to\u003c/p\u003e\u003cp\u003edomain-specific optimization, such as Huawei Cloud\u0026rsquo;s MindBot, which is dedicated to Chinese mental health services. These advances provide a foundation for our Dual-Channel evaluation model.\u003c/p\u003e\u003c/div\u003e"},{"header":"3 Methodology","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n\u003ch2\u003e3.1 Selection of the Base Model\u003c/h2\u003e\n\u003cp\u003eCurrently, mainstream large language models mainly include ChatGPT, DeepSeek, Google Gemini, QwenLM, etc. Referring to the research by Rahman et al. [\u003cspan class=\"CitationRef\"\u003e36\u003c/span\u003e] and combining the characteristics of anx- iety risk assessment tasks, this paper selects the following metrics to compare the performance of each model:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003eMMLU (Overall): Using the Massive Multitask Language Understanding (MMLU) benchmark, it assesses factual recall, contextual reasoning, and complex question-answering capabilities across diverse knowledge domains.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eMMLU\u0026ndash;Reasoning: The MMLU-Reasoning Subset evaluates a model\u0026rsquo;s ability to process multi-step, context-dependent questions within specialized domains like law or biology.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eHellaSwag: HellaSwag benchmarks commonsense reasoning by testing a model\u0026rsquo;s ability to predict the most plausible continuation of everyday scenarios. It measures intuitive understanding of physical and social contexts.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eMATH-500: The MATH-500 dataset assesses numerical and quantitative reasoning by requiring models to solve structured mathematical problems, including symbolic manipulation and equation solving. Scores reflect logical deduction efficiency.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eUSMLE Step 1: The USMLE Step 1 benchmark evaluates medical knowledge integration, cover- ing pathophysiology, pharmacology, and diagnostics. Models are tested on clinical decision-making accuracy.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eCommonsenseQA: CommonsenseQA tests everyday reasoning through multiple-choice questions requiring implicit world knowledge. It highlights gaps in models\u0026rsquo; real-world semantic understanding.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe final results are shown in Figure.1. It can be found that for medical texts, the DeepSeek model has the best comprehensive performance with the least resource consumption. Therefore, this paper selects the DeepSeek-R1 model as the base model.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n\u003ch2\u003e3.2 Model Training and Optimization\u003c/h2\u003e\n\u003cp\u003eWe used DeepSeek-R1-Distill-Llama-8B as the base model and adopted 4-bit quantization (GPTQ algorithm) through the Unsloth optimization framework, achieving efficient operation with 4096\u003c/p\u003e\n\u003cp\u003etokens maximum sequence length. The system implements dynamic conversation history management (maximum 8 dialogue rounds) with automatic context pruning to maintain optimal performance.\u003c/p\u003e\n\u003cp\u003eKey hyperparameters were set as follows:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003eGeneration parameters: Primary response: temperature\u0026thinsp;=\u0026thinsp;0.7; top-p\u0026thinsp;=\u0026thinsp;0.9; repetition penalty\u0026thinsp;=\u0026thinsp;1.1; Emotion analysis: temperature\u0026thinsp;=\u0026thinsp;0.3 (deterministic sampling); Max new tokens: 800 for responses, 50 for emotion classification.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eMemory optimization: 4-bit quantization reducing memory demand to 15GB total utilization; Automatic history truncation preserving 16 most recent exchanges.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eArchitecture features: Two-stage generation (emotion classification \u0026rarr; contextual response); Explicit clinical reasoning via \u0026lt;\u0026thinsp;think\u0026thinsp;\u0026gt;\u0026thinsp;tags; Role-specific formatting (\u0026ldquo;client\u0026rdquo;/\u0026ldquo;Dr.Ivy\u0026rdquo;).\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe dialogue system demonstrates: real-time interaction speed maintained at \u0026lt;\u0026thinsp;2s latency per round; stable context utilization through 4096-token window management; clinical response quality via pre- response emotion analysis, structured reasoning prompts, and multi-round context preservation.\u003c/p\u003e\n\u003cp\u003eFine-tuning preserved the original 12.5% memory footprint of FP32 while extending capabilities for psychological counseling applications. The system shows particular effectiveness in maintaining natural dialogue flow despite reduced repetition penalty, delivering clinically-structured responses via template engineering, and balancing response diversity (top-p sampling) with consistency (temperature tuning).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n\u003ch2\u003e3.3 Dual-Channel Risk Assessment Model\u003c/h2\u003e\n\u003cp\u003eWe adopted an innovative Dual-Channel BERT architecture for anxiety risk assessment. A shared BERT- base feature extraction layer was used to obtain text semantic representations (hidden state H \u0026isin; Rn \u0026times;768 ), with parallel operation of the main classification channel and high-risk detection channel.The main classification channel outputs probability distributions of low, moderate, and high anxiety through fully connected layers, as expressed in (1):\u003c/p\u003e\n\u003cp\u003ePmain\u0026thinsp;=\u0026thinsp;Softmax (Wm \u0026middot; Dropout (h[CLS])\u0026thinsp;+\u0026thinsp;bm ) (1)\u003c/p\u003e\n\u003cp\u003eThe Dual-Channel BERT architecture we designed is shown in Fig.\u0026nbsp;2. The high-risk channel uses a deep compression structure (768\u0026rarr; 64\u0026rarr; 1) and a Sigmoid activation function to enhance the identification of high-risk features, as shown in (2):\u003c/p\u003e\n\u003cp\u003ePhigh\u0026thinsp;=\u0026thinsp;\u0026sigma; (W2 \u0026middot; ReLU (W1 \u0026middot; h[CLS] )) (2)\u003c/p\u003e\n\u003cp\u003eTo address annotation noise, we designed a dynamic label smoothing strategy: 30% of high-risk boundary samples were re-annotated as moderate risk, and a weighting coefficient \u0026alpha;\u0026thinsp;=\u0026thinsp;0.7 was applied to the loss function, as shown in (3):\u003c/p\u003e\n\u003cp\u003ewhere CE(\u0026middot;) denotes the cross-entropy loss, Ai is the annotation confidence, and yi is the true label.We used a Dual-Channel mechanism in the decision-making stage: If the highest class probability in the main channel was \u0026lt;\u0026thinsp;0.7 and the high-risk channel output exceeded 0.4, the sample was classified as high-risk; otherwise, the main channel result was retained. A weighted loss function (with \u0026lambda;\u0026thinsp;=\u0026thinsp;2.0) was used to balance overall classification performance and high-risk recall, as shown in (4):\u003c/p\u003e\n\u003cp\u003eLtotal\u0026thinsp;=\u0026thinsp;Lmain\u0026thinsp;+\u0026thinsp;\u0026lambda; \u0026middot; BCE(Phigh, I[y\u0026thinsp;=\u0026thinsp;2]) (4)\u003c/p\u003e\n\u003cp\u003ewhere BCE(\u0026middot;) denotes the binary cross-entropy loss, and I[\u0026middot;] is an indicator function (1 if y\u0026thinsp;=\u0026thinsp;2 (high risk), 0 otherwise). The model achieved a macro F1-score of 0.8886 on the test set, with an 3% increase in high-risk recall compared with the single-channel baseline.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n\u003ch2\u003e3.4 System Integration and Real-Time Interaction\u003c/h2\u003e\n\u003cp\u003eWe integrated the fine-tuned LLM with a Dual-Channel evaluation module to form an end-to-end interac- tive framework. During conversations, user inputs were encoded by BERT and processed simultaneously by the psychological support layer (generating CBT-based responses) and the risk assessment layer (cal- culating anxiety risks). Real-time risk assessment feedback was used to adjust the dialogue engine\u0026rsquo;s intervention strategies.\u003c/p\u003e\n\u003cdiv class=\"BlockQuote\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003cp\u003eThe system had an average processing time of \u0026lt;\u0026thinsp;1 second per round and supported context tracking in over 15 rounds of extended dialogues. Its dynamic attention mechanism captured temporal behavioral patterns (e.g., emotional spikes or risk escalations). The output included natural language responses and visualized risk levels (low/moderate/high), with a clinical false positive rate\u0026thinsp;\u0026lt;\u0026thinsp;5% (below the safety threshold).\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4 Experiment","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n\u003ch2\u003e4.1 Datasets and Annotation Methods\u003c/h2\u003e\n\u003cp\u003eWe used two key datasets for model training and validation:\u003c/p\u003e\n\u003cp\u003e1) The aiwei.json dataset: The aiwei.json dataset from the EmoLLM project (900 psychological coun- seling dialogues) was used for LLM fine-tuning. This dataset simulates professional counseling scenarios, with system prompts defining the model role as \u0026ldquo;Dr. Aiwei, a gentle senior nurse with rich psycho- logical expertise.\u0026rdquo; Each data entry includes user-reported psychological issues (input) and professional responses compliant with cognitive behavioral therapy (CBT) protocols (output), optimized using a thought chain template: \u0026ldquo;Step 1: Emotional recognition \u0026rarr; Step 2: Empathetic expression \u0026rarr; Step 3: Solution suggestions \u0026rarr; Step 4: Safety verification.\u0026rdquo;\u003c/p\u003e\n\u003cp\u003e2) SoulChatCorpus-sft-multi-turn mixed dataset: This dataset is used to validate anxiety risk assess- ment. It contains 150,000 single-round consultations and 1\u0026nbsp;million multi-round conversations, with high-quality samples retained after rigorous privacy filtering. To address the lack of annotations, we developed a four-level classification system that strictly aligns with the clinical frameworks of the GAD-7 and HAMA scales, encompassing four dimensions: core symptoms, somatic symptoms, cognitive symp- toms, and functional impairment. The categories are defined as follows: No significant anxiety (using vague terms like \u0026rdquo;somewhat,\u0026rdquo; corresponding to GAD-7 scores of 1\u0026ndash;4); Mild anxiety (using absolute terms like \u0026rdquo;always,\u0026rdquo; corresponding to GAD-7 scores of 5\u0026ndash;9); Moderate anxiety (using explicit negative emotion\u003c/p\u003e\n\u003cp\u003eexpressions, corresponding to GAD-7 scores of 10\u0026ndash;14); Severe anxiety (characterized by extreme expres- sions or severe somatic symptoms, corresponding to GAD-7 scores of 15\u0026ndash;21). A team of 10 annotators underwent standardized training in GAD-7 criteria, conducted multiple rounds of annotation on the orig- inal samples, and implemented a dual-review mechanism. To verify annotation reliability, inter-annotator consistency was assessed using 500 randomly selected samples, yielding a Fleiss\u0026rsquo; Kappa value of 0.82, indicating excellent annotation consistency. To address category imbalance in the original dataset, a core dataset was constructed based on clinical annotation standards. Targeted data augmentation strategies were then applied to generate final training/validation/testing data, including synonym substitution, random deletion (5%-15% non-keywords), and sentence restructuring. This produced 1,530 augmented samples for moderate anxiety and 1,707 for severe anxiety. The enhanced dataset ultimately contained 15,078 user messages with balanced category distribution: 5,041 for no significant anxiety, 6,023 for mild anxiety, 2,007 for moderate anxiety, and 2,007 for severe anxiety. A stratified random partitioning strat- egy was then employed to ensure balanced distribution across training, validation, and testing sets in a 14:3:3 ratio.\u003c/p\u003e\n\u003cp\u003e3)The generalization capability was validated using an external PsyDTCorpus dataset with identical annotation protocols, following the same labeling approach as the SoulChatCorpus-sft-multi-turn dataset. Inter-rater reliability was assessed across 500 randomly selected samples, yielding a Fleiss\u0026rsquo; Kappa score of 0.85, demonstrating excellent annotation consistency.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\n\u003ch2\u003e4.2 Innovative Framework\u003c/h2\u003e\n\u003cp\u003eWe proposed a \u0026ldquo;dynamic interaction\u0026thinsp;+\u0026thinsp;hierarchical evaluation\u0026rdquo; anxiety risk analysis paradigm, with the system architecture shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\n\u003cdiv class=\"BlockQuote\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003ePsychological Support Layer: A psychological counseling module using Chain-of-Thought (CoT) rea- soning decomposed traditional CBT into interpretable reasoning steps. It included an emotion state tracker to maintain the user\u0026rsquo;s real-time emotional vector space.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u0026nbsp;Risk Assessment Layer: A Dual-Channel BERT method was used to improve the BERT BASE CHINESE model. It extracted anxiety semantics through a main classification channel, a high-risk clas- sification channel, and a fuzzy labeling method, with anxiety risk signals fed back to the psychological support layer.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n\u003ch2\u003e4.3 Model Fine-tuning Performance\u003c/h2\u003e\n\u003cp\u003eWe use Unsloth optimization framework to implement 4bit quantization training, and the key parameters of training are shown in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003ctable id=\"Tab1\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eTraining Parameters and Technical Advantages\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eParameter\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eValue\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eTechnical Benefit\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eQuantization Method\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eGPTQ 4bit\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e62% VRAM reduction with maintained accuracy\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eBatch Size\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e4\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e3.2 \u0026times; throughput improvement\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eGradient Accumulation Steps\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e8\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eEnables large virtual batches\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLearning Rate\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e3 \u0026times; 10\u0026thinsp;\u0026minus;\u0026thinsp;5\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eOptimized for 8B-scale convergence\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eOptimizer\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eAdamW-bnb-8bit\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eGradient quantization error\u0026thinsp;\u0026lt;\u0026thinsp;0.8%\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eThe fine-tuned DeepSeek-R1-Distill-Llama-8B demonstrated superior efficiency, achieving 72 token- s/s inference speed with 15GB VRAM usage (4-bit quantized) while maintaining a 4096-token context\u003c/p\u003e\n\u003cp\u003ewindow - outperforming comparable models like ChatGLM-6B (42 tokens/s, 13.2GB, 2048-token) and Qwen-1.8B (68 tokens/s, 5.8GB, 2048-token). As evidenced in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e, our optimized training protocol maintained exceptional gradient stability, with norms converging to 0.51\u0026thinsp;\u0026plusmn;\u0026thinsp;0.01 after 60 steps (vs. typ- ical 0.5-2.0 benchmarks), while Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e shows the loss descending from 2.1 to 1.1 (48% reduction) with logarithmic consistency (R2\u0026thinsp;=\u0026thinsp;0.97).\u003c/p\u003e\n\u003cp\u003eThis represents a 1.7 \u0026times; speedup over baseline LLaMA2-7B (55 tokens/s, 18GB) despite our model\u0026rsquo;s larger context capacity, attributable to: 1) dynamic gradient scaling preventing norm fluctuations beyond the 0.5\u0026ndash;0.75 range, and 2) Unsloth\u0026rsquo;s block-wise recomputation enabling stable 4-bit training.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\n\u003ch2\u003e4.4 Anxiety Risk Assessment Performance\u003c/h2\u003e\n\u003cp\u003eThis study employs five leading models\u0026mdash;BERT BASE CHINESE, MentalBERT, GLM-4-Air, CNN, and LSTM\u0026mdash;as baselines. The proposed Dual-Channel BERT was tested against these models on the unified SoulChatCorpus-sft-multi-turn dataset for anxiety risk assessment. The classification performance differences are detailed in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e and Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\n\u003cdiv id=\"Sec16\" class=\"Section3\"\u003e\n\u003ch2\u003e4.4.1 Classification Result\u003c/h2\u003e\n\u003cp\u003e\u0026bull; BERT BASE CHINESE: The overall accuracy rate is 86.38%, and the weighted F1 score is also 86.38%. It performs relatively stable in the categories of no significant anxiety and mild anxiety. However, the\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003cdiv class=\"colspec\" align=\"char\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003ctable id=\"Tab2\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eClassification Result\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eModel Type\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eTrue\u003c/p\u003e\n\u003cp\u003eLabel\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003ePredicted\u003c/p\u003e\n\u003cp\u003eas none\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003ePredicted\u003c/p\u003e\n\u003cp\u003eas low\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003ePredicted\u003c/p\u003e\n\u003cp\u003eas mid\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003ePredicted\u003c/p\u003e\n\u003cp\u003eas high\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd rowspan=\"4\" align=\"left\"\u003e\n\u003cp\u003eBERT\u003c/p\u003e\n\u003cp\u003eBASE CHINESE\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003enone\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e615\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e137\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e4\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003elow\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e63\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e810\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e25\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e6\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003emid\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e48\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e242\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e11\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ehigh\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e2\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e12\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e287\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd rowspan=\"4\" align=\"left\"\u003e\n\u003cp\u003eMental\u003c/p\u003e\n\u003cp\u003eBERT\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003enone\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e650\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e99\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e5\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e2\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003elow\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e84\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e780\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e35\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e5\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003emid\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e27\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e255\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e19\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ehigh\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e18\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e283\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd rowspan=\"4\" align=\"left\"\u003e\n\u003cp\u003eGLM-4\u003c/p\u003e\n\u003cp\u003e-Air\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003enone\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e583\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e169\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e4\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003elow\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e48\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e574\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e277\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e5\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003emid\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e275\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e24\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ehigh\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e5\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e79\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e217\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd rowspan=\"4\" align=\"left\"\u003e\n\u003cp\u003eCNN\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003enone\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e559\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e195\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003elow\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e73\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e811\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e18\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e2\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003emid\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e31\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e270\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ehigh\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e2\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e13\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e285\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd rowspan=\"4\" align=\"left\"\u003e\n\u003cp\u003eLSTM\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003enone\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e586\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e166\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e3\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003elow\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e93\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e770\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e37\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e4\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003emid\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e31\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e268\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e2\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ehigh\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e5\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e3\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e293\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd rowspan=\"4\" align=\"left\"\u003e\n\u003cp\u003eDual-Channel\u003c/p\u003e\n\u003cp\u003eBERT\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003enone\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e640\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e112\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e3\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003elow\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e87\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e796\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e20\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003emid\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e25\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e275\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ehigh\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e2\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e299\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003cdiv class=\"colspec\" align=\"char\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003ctable id=\"Tab3\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eCompare Experimental Results-SoulChatCorpus\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eModel Type\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003ePrecision\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eRecall\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eF1\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eAccuracy\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eBERT BASE CHINESE\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8673\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8638\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8638\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8638\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMentalBERT\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8700\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8703\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8700\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8700\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eGLM-4-Air\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.7892\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.7290\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.7419\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.7290\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eCNN\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8589\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8510\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8509\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8510\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLSTM\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8495\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8475\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8473\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8475\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003eDual-Channel BERT\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.8886\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.8889\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.8886\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.8886\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003cdiv class=\"colspec\" align=\"char\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003ctable id=\"Tab4\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eCompare Experimental Results-PsyDTCorpus\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eModel Type\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003ePrecision\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eRecall\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eF1\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eAccuracy\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eBERT BASE CHINESE\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.9175\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.9145\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.9153\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.9145\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMentalBERT\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.9271\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.9273\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.9272\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.9273\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eGLM-4-Air\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.7882\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.7259\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.7392\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.7259\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eCNN\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8961\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8966\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8962\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8966\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLSTM\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8976\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8972\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8973\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8972\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003eDual-Channel BERT\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.9319\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.9323\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.9320\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.9323\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003erecall rate for moderate anxiety is only 80.40%, while the recall rate for severe anxiety reaches 95.35%, though there are still a few high-risk samples that were missed.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003eMentalBERT: Its overall performance is comparable to BERT BASE CHINESE, achieving an accuracy rate of 87.00% and a weighted F1 score of 87.00%. It demonstrates good consistency in classifying no anxiety and mild anxiety, but falls slightly short of BERT BASE CHINESE in identifying moderate anxiety (81.47%) and severe anxiety (91.59%) with higher-risk categories.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eGLM-4-Air: The overall performance was the poorest,with an accuracy rate of only 72.90% and a weighted F1 score of 74.19%. There was significant category confusion, as mild anxiety was misclassified as moderate anxiety in 277 cases. The recall rate for severe anxiety was only 72.09%, failing to meet the reliability requirements for clinical screening.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eCNN achieved an overall accuracy of 85.10% and a weighted F1 score of 85.09%. The system demon- strated strong performance in classifying moderate and severe anxiety (with a F1 score of 96.77% for severe anxiety), but its recall rate for anxiety was only 73.94%, indicating a lack of distinct anxiety categories. Additionally, low-risk samples were frequently misclassified as mild anxiety.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eLSTM: Overall accuracy of 84.75% and weighted F1 score of 84.73%. It demonstrates outstanding per- formance in severe anxiety recognition (F1 score: 97.18%), but shows no significant confusion between anxiety and mild anxiety categories, with generally average classification stability.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eThe Dual-Channel BERT model achieves optimal overall performance with an accuracy of 88.86% and a weighted F1 score of 88.86%. It demonstrates superior balance and reliability in anxiety category classification compared to the baseline model, particularly with a 99.34% recall rate for severe anxiety cases. The model exhibits an extremely low false negative rate for high-risk samples, fully meeting clinical intervention priorities.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec17\" class=\"Section3\"\u003e\n\u003ch2\u003e4.4.2 Ablation Studies\u003c/h2\u003e\n\u003cp\u003eTo evaluate the performance gains of the dual-channel architecture and fuzzy tag processing mechanism, an ablation experiment was conducted. Using the SoulChatCorpus-sft-multi-turn dataset, the study compared four models: the full Dual Channels BERT model, the model with removed fuzzy tags, the model with removed dual channels, and the model with both removed. The results are presented in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003eWhen fuzzy labels are removed, the model\u0026rsquo;s precision, recall, F1 score, and accuracy all drop to 0.8703\u0026ndash;0.8700. This demonstrates that the fuzzy label processing mechanism effectively mitigates category boundary ambiguity and enhances the model\u0026rsquo;s ability to capture semantic nuances. The absence of this mechanism would otherwise lead to reduced classification stability.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eDisabling dual channels (Remove Dual Channels) results in a performance drop (0.8783\u0026ndash;0.8784), demonstrating that the dual-channel architecture optimizes feature extraction and classification logic through coordinated decision-making between the primary classification channel and high-risk channel. A single channel cannot adequately capture the complex semantic features of anxiety risk.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eSimultaneously removing both (Remove Dual Channels and Fuzzy Tags) reduced performance to 0.8673\u0026ndash;0.8638, demonstrating the synergy between dual-channel and fuzzy-label mechanisms: fuzzy labels optimize annotation quality while dual channels enhance feature utilization, jointly ensuring the model\u0026rsquo;s accurate identification of varying anxiety risk levels.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eThe full model (Dual Channels BERT) achieved precision, recall, F1 score, and accuracy of 0.8889\u0026ndash;0.8886, significantly outperforming the post-ablation model. This demonstrates that the dual- channel architecture and fuzzy label processing mechanism are key design elements for enhancing anxiety risk assessment performance, effectively addressing core issues such as category confusion and high-risk misclassification.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tab5\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eAblation Experiment Results\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eModel Type\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003ePrecision\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eRecall\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eF1\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eAccuracy\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eRemove Fuzzy Tags\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8703\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8700\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8700\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8703\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eRemove Dual Channels\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8783\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8784\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8783\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8784\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eRemove Dual Channels and Fuzzy Tags\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8673\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8638\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8638\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.8638\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003eDual Channels BERT\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.8889\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.8886\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.8886\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e\u003cstrong\u003e0.8886\u003c/strong\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003c/div\u003e\n\u003c/div\u003e"},{"header":"5 A Real-world Case Study","content":"\u003cp\u003eTo further validate our algorithm, we designed a user-friendly interactive interface that facilitates seam- less therapeutic interaction while providing real world anxiety monitoring. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e6\u003c/span\u003e, the interface adopts a dual-panel layout: the left panel hosts the conversational chatbot where users interact with virtual therapist \u0026rdquo;Dr. Ava,\u0026rdquo; while the right panel displays analytical components including real- time anxiety scores, a dynamic trend chart tracking emotional fluctuations across conversation turns, and a comprehensive score distribution histogram. The interface incorporates a status indicator show- ing system connection status, along with a dedicated function for generating automated anxiety analysis reports summarizing key statistics (average score, maximum/minimum values, standard deviation) and\u003c/p\u003e\u003cp\u003etrend interpretation. This design enables simultaneous therapeutic dialogue and quantitative mental state assessment, creating an immersive environment that balances clinical functionality with user-friendly interaction.\u003c/p\u003e\u003cp\u003eTo validate the practical effectiveness of our anxiety assessment system, we conducted a real-time case study using the AI Psychotherapy System interface Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e7\u003c/span\u003e The system monitored a user initially expressing significant social anxiety regarding friendship loss and difficulties in making new connections.\u003c/p\u003e\u003cp\u003eDuring the 8-round therapeutic dialogue, the user first revealed: \u0026rdquo;I\u0026rsquo;m afraid of losing existing friend- ships and feel hesitant to actively make new friends.\u0026rdquo; Through structured guidance from virtual therapist \u0026rdquo;Dr. Ava\u0026rdquo;, the conversation progressively addressed these concerns. The system\u0026rsquo;s dual-channel archi- tecture successfully detected subtle linguistic markers of anxiety while maintaining natural therapeutic interaction. As shown in the Anxiety Score Trend chart, the initial anxiety score reached 0.44 (round 2) when the user expressed core fears, but gradually decreased to 0.268 by round 8 after receiving cognitive restructuring and behavioral suggestions.\u003c/p\u003e\u003cp\u003eThe Score Distribution histogram demonstrates that 62.5% of conversation rounds maintained scores below 0.4, with only one round exceeding 0.44, indicating effective anxiety regulation during the ses- sion. This case vividly demonstrates our system\u0026rsquo;s comprehensive capability to first identify acute anxiety triggers in real-time through linguistic pattern recognition, then deliver appropriate therapeutic inter- ventions while continuously monitoring emotional fluctuations through its dual-channel architecture, and finally document the complete anxiety resolution journey from initial problem exposure to successful cog- nitive restructuring. The interface not only effectively supported the entire therapeutic process but also provided quantifiable anxiety metrics that tracked the user\u0026rsquo;s emotional progression, ultimately facilitat- ing a successful transition from avoidance behavior characterized by hesitation in making new friends to adopting proactive solutions such as joining tour groups for social connection.\u003c/p\u003e\u003cp\u003eWhile the presented case demonstrates success, our system still has limitations including potential underestimation of intervention effects and risk assessment errors. When the system detects three con- secutive rounds of no significant reduction in user anxiety scores (fluctuation i0.05), it triggers a warning mechanism advising users: \u0026rdquo;This issue may require professional face-to-face intervention. We recommend contacting local mental health services.\u0026rdquo; The system\u0026rsquo;s built-in error detection module automatically flags cases where high-risk assessments show below-threshold consistency with dialogue context (0.6), with reports noting \u0026rdquo;This risk assessment requires further confirmation by professionals.\u0026rdquo; Example failure: A user describing \u0026rdquo;recent frequent insomnia without noticeable impact on work or life\u0026rdquo; was misclassified as moderate anxiety (actual label: mild anxiety), reflecting the model\u0026rsquo;s excessive weighting of somatic symptoms. Future iterations will optimize the correlation modeling between symptoms and functional impacts.\u003c/p\u003e"},{"header":"6 Conclusion","content":"\u003cp\u003eWhile traditional psychological anxiety risk assessment methods demonstrate good reliability and valid- ity, they require substantial clinical expertise from practitioners, limiting their scalability. These methods also face risks of patients deliberately concealing their conditions. Moreover, conventional approaches struggle to capture dynamic behavioral data or reveal potential correlations between anxiety states, ver- bal expressions, and interactive behaviors. This study develops a large-model-based psychological anxiety risk assessment system that evaluates users\u0026rsquo; anxiety risks through monitoring their interactions with the model. The research focuses on three key aspects: Methodologically, we propose a \u0026rdquo;dynamic interac- tion\u0026thinsp;+\u0026thinsp;hierarchical assessment\u0026rdquo; paradigm, with annotation standards strictly aligned to the GAD-7 and HAMA clinical scales, overcoming the static limitations of traditional assessments. Through LoRA fine- tuning and collaborative optimization of four-dimensional quantification, the system maintains clinical validity while reducing resource consumption by 87%, providing a solution for deployment in resource- constrained scenarios. Technically, the dual-channel BERT model significantly enhances high-risk sample identification through fuzzy label processing and collaborative decision-making mechanisms. Experiments show a 7.1% improvement in robustness under noisy labeling conditions and a 99.34% high-risk recall rate. In terms of application value, the end-to-end system achieves real-time assessment capability of processing 8.2 dialogue rounds per minute, supporting dynamic tracking of anxiety risks. By designing ethical intervention mechanisms and clarifying clinical assistance positioning, the system reduces misuse risks in high-risk scenarios. By encoding clinical criteria into an interpretable dialogue process (PRDS strategy), this model ensures interventions comply with CBT treatment protocols. Experimental results demonstrate that the proposed model exhibits high accuracy and robustness in anxiety risk assessment,\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eeffectively complementing traditional scales and clinical interviews while providing new technical sup- port for digital mental health interventions. The system is designed for risk assessment and early support rather than crisis intervention. When detecting high-risk expressions like suicidal ideation, it offers empa- thetic support while strongly recommending professional help and stopping further discussion to prevent potential harm. System performance is limited by training data and lacks actual emergency response capabilities. All assessment results should be used as clinical references and cannot replace professional diagnosis.\u003c/p\u003e\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e\u003ch2\u003e6.1 Future Work\u003c/h2\u003e\u003cp\u003eDespite the encouraging outcomes of this study in psychological anxiety risk identification, several limi- tations should be acknowledged, pointing toward meaningful directions for future research. One primary area for improvement lies in the enhancement and diversification of training data. The current dataset may not adequately represent all demographic groups or mental health conditions, which constrains the model\u0026rsquo;s interpretability and robustness. Subsequent efforts would benefit from incorporating more diverse data sources\u0026mdash;such as social media content, news articles, and medical platform records\u0026mdash;to improve generalizability and practical applicability.\u003c/p\u003e\u003cp\u003e Furthermore, the reliance solely on textual data limits the model\u0026rsquo;s ability to capture non-verbal behavioral cues. Developing a multimodal evaluation framework that integrates audio, visual, and other forms of data could significantly enhance detection accuracy by offering a more comprehensive view of user expressions.\u003c/p\u003e\u003cp\u003eNext, we will conduct prospective clinical validation studies through parallel assessments with psy- chiatrists to further validate the diagnostic accuracy, reliability, and practicality of the clinical utility verification model in real-world clinical settings. This will clarify its applicability as an auxiliary screening tool, establish operational protocols and quality control standards, and promote the seamless integration of digital assessment tools with traditional psychiatric diagnostic workflows.\u003c/p\u003e\u003cp\u003eAnother promising direction involves extending the model\u0026rsquo;s capability to multiple languages and cultural contexts. While the current system performs effectively in Chinese settings, its performance degrades when applied to other languages. Given the global and culturally varied nature of mental health needs, future work should prioritize training on multilingual and culturally diverse datasets. Such efforts would foster greater inclusivity and improve the model\u0026rsquo;s sensitivity to linguistic and cultural nuances, enabling more reliable anxiety risk detection across different populations.\u003c/p\u003e\u003cp\u003eIn conclusion, the anxiety risk identification model developed in this study provides a useful supple- mentary tool for round-the-clock mental health monitoring through conversational interactions. It holds promise as a scalable aid to existing clinical screening practices. Nonetheless, continued innovation and refinement are essential to broaden its applicability and strengthen its real-world impact.\u003c/p\u003e\u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics\u0026nbsp;approval\u0026nbsp;and\u0026nbsp;consent\u0026nbsp;to\u0026nbsp;participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot\u0026nbsp;applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent\u0026nbsp;for\u0026nbsp;publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot\u0026nbsp;applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability\u003c/strong\u003e\u003cstrong\u003eof data\u003c/strong\u003e\u003cstrong\u003eand\u003c/strong\u003e\u003cstrong\u003ematerial\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe \u0026nbsp;datasets \u0026nbsp; analysed \u0026nbsp;during \u0026nbsp; the \u0026nbsp;current \u0026nbsp;study\u0026nbsp;\u0026nbsp;are\u0026nbsp;\u0026nbsp;available\u0026nbsp;\u0026nbsp;in \u0026nbsp;the\u0026nbsp;\u0026nbsp;EmoLLM\u0026nbsp;\u0026nbsp;(aiwei.json)\u0026nbsp;\u0026nbsp;and SoulChatCorpus\u0026nbsp;repositories\u0026nbsp;on\u0026nbsp;GitHub,\u0026nbsp;accessible\u0026nbsp;at:\u003c/p\u003e\n\u003cp\u003e•\u0026nbsp;https://github.com/SmartFlowAI/EmoLLM/blob/main/datasets/aiwei.json\u003c/p\u003e\n\u003cp\u003e•\u0026nbsp;https://github.com/scutcyr/SoulChat\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting\u003c/strong\u003e\u003cstrong\u003einterests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe\u0026nbsp;authors\u0026nbsp;declare\u0026nbsp;that\u0026nbsp;they\u0026nbsp;have\u0026nbsp;no\u0026nbsp;competing\u0026nbsp;interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot\u0026nbsp;applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors’\u003c/strong\u003e\u003cstrong\u003econtributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eY. L. J. and H. Z. Y. were responsible for the writing of the original draft; L. L. L., D. Z. Y. and Z. Y. critically reviewed and edited the manuscript; Z. Y. served as the corresponding author, handling all correspondence and coordination with the journal during submission and peer review. All authors have read and approved the final submitted version and agree to be accountable for all aspects of the work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eHirsch CR, Mathews A. A cognitive model of pathological worry. Behav Res Ther. 2012;50(10):636\u0026ndash;46.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChand SP, Marwaha R, Bender RM. Anxiety (nursing) (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSpitzer RL, Kroenke K, Williams JB, Lo\u0026hellip; we B. A brief measure for assessing generalized anxiety disorder: the gad-7. Arch Intern Med. 2006;166(10):1092\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePlummer F, Manea L, Trepel D, McMillan D. Screening for anxiety disorders with the gad-7 and gad-2: a systematic review and diagnostic metaanalysis. Gen Hosp Psychiatry. 2016;39:24\u0026ndash;31.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTorous J, Myrick KJ, Rauseo-Ricupero N, Firth J, et al. Digital mental health and covid-19: using technology today to accelerate the curve on access and quality tomorrow. JMIR mental health. 2020;7(3):18848.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWeizenbaum J. Eliza\u0026mdash;a computer program for the study of natural language communication between man and machine. Commun ACM. 1966;9(1):36\u0026ndash;45.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOrru G, Pettersson-Yeo W, Marquand AF, Sartori G, Mechelli A. Using support vector machine to identify imaging biomarkers of neurological and psychiatric disease: a critical review. Neurosci Biobehavioral Reviews. 2012;36(4):1140\u0026ndash;52.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDe Choudhury M, Gamon M, Counts S, Horvitz E. Predicting depression via social media. In: Proceedings of the International AAAI Conference on Web and Social Media, vol. 7, pp. 128\u0026ndash;137 (2013).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhao W, Cai Z, Dong S, Hei W. A review of machine learning-based assessment of depression. In: China Health Information Processing Conference, pp. 266\u0026ndash;290 (2024). Springer.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKhoo LS, Lim MK, Chong CY, McNaney R. Machine learning for multimodal mental health detection: a systematic review of passive sensing approaches. Sensors. 2024;24(2):348.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNaveed H, Khan AU, Qiu S, Saqib M, Anwar S, Usman M, Akhtar N, Barnes N, Mian A. A comprehensive overview of large language models. ACM Transactions on Intelligent Systems and Technology; 2023.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDevlin J, Chang M-W, Lee K, Toutanova K: Bert: Pre-training of deep bidirectional trans- formers for language understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (long and, Papers S.), pp. 4171\u0026ndash;4186 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePampouchidou A, Pediaditis M, Kazantzaki E, Sfakianakis S, Apostolaki I-A, Argyraki K, Manousos D, Meriaudeau F, Marias K, Yang F, et al. Automated facial video-based recog- nition of depression and anxiety symptom severity: cross-corpus validation. Mach Vis Appl. 2020;31(4):30.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDiep B, Stanojevic M, Novikova J. Multi-modal deep learning system for depression and anxiety detection. arXiv preprint arXiv:221214490 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDanner M, Hadzic B, Gerhardt S, Ludwig S, Uslu I, Shao P, Weber T, Shiban Y, Ratsch M. Advancing mental health diagnostics: Gpt-based method for depression detection. In: 2023 62nd Annual Conference of the Society of Instrument and Control Engineers (SICE), pp. 1290\u0026ndash;1296 (2023). IEEE.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGuo Z, Lai A, Thygesen JH, Farrington J, Keen T, Li K, et al. Large language models for mental health applications: systematic review. JMIR mental health. 2024;11(1):57400.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePerlis RH, Goldberg JF, Ostacher MJ, Schneck CD. Clinical decision support for bipolar depression using large language models. Neuropsychopharmacology. 2024;49(9):1412\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHamilton M. The assessment of anxiety states by rating. Br J Med Psychol (1959).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZigmond AS, Snaith RP. The hospital anxiety and depression scale. Acta psychiatrica scandi- navica. 1983;67(6):361\u0026ndash;70.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eStafford L, Berk M, Jackson HJ. Validity of the hospital anxiety and depression scale and patient health questionnaire-9 to screen for depression in patients with coronary artery disease. Gen Hosp Psychiatry. 2007;29(5):417\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHasin DS, Greenstein E, Aivadyan C, Stohl M, Aharonovich E, Saha T, Goldstein R, Nunes EV, Jung J, Zhang H, et al. The alcohol use disorder and associated disabilities interview schedule-5 (audadis-5): procedural validity of substance use disorders modules through clinical re- appraisal in a general population sample. Drug Alcohol Depend. 2015;148:40\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKaufman J, Kobak K, Birmaher B, Lacy N. Ksads-comp perspectives on child psychiatric diagnostic assessment and treatment planning. J Am Acad Child Adolesc Psychiatry. 2020;60(5):540\u0026ndash;2.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePestian JP, Matykiewicz P, Linn-Gust M, South B, Uzuner O, Wiebe J, Cohen KB, Hurdle J, Brew C. Sentiment analysis of suicide notes: A shared task. Biomedical Inf insights. 2012;5:9042.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDe Choudhury M, Counts S, Horvitz EJ, Hoff A. Characterizing and predicting postpartum depression from shared facebook data. In: Proceedings of the 17th ACM Conference on Computer Supported Cooperative Work \u0026amp; Social Computing, pp. 626\u0026ndash;638 (2014).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi D, Niu J, Qiu M, Liu M. Sentiment analysis on weibo data. In: 2014 IEEE Computers, Communications and IT Applications Conference, pp. 249\u0026ndash;254 (2014). IEEE.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIlikci B, Chen L, Cho H, Liu Q. Heat-map based emotion and face recognition from thermal images. In: 2019 Computing, Communications and IoT Applications (ComComAp), pp. 449\u0026ndash;453 (2019). IEEE.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTadesse MM, Lin H, Xu B, Yang L. Detection of depression-related posts in reddit social media forum. IEEE Access. 2019;7:44883\u0026ndash;93.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGuntuku SC, Yaden DB, Kern ML, Ungar LH, Eichstaedt JC. Detecting depression and mental illness on social media: an integrative review. Curr Opin Behav Sci. 2017;18:43\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser L\u0026ndash;. Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJi S, Zhang T, Ansari L, Fu J, Tiwari P, Cambria E. Mentalbert: Publicly available pretrained language models for mental healthcare. arXiv preprint arXiv:211015621 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFitzpatrick KK, Darcy A, Vierhile M. Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (woebot): a randomized controlled trial. JMIR mental health. 2017;4(2):7785.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang X, Sharma D, Kumar D. A review on ai-based modeling of empathetic conversational response generation. In: 2023 Asia Conference on Cognitive Engineering and Intelligent Interaction (CEII), pp. 102\u0026ndash;109 (2023). IEEE.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGrabb D. The impact of prompt engineering in large language model performance: a psychiatric example. J Med Artif Intell 6 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang W, Deng Y, Liu B, Pan SJ, Bing L. Sentiment analysis in the era of large language models: A reality check. arXiv preprint arXiv:230515005 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDing Y, Liu J, Zhang X, Yang Z. Dynamic tracking of state anxiety via multi-modal data and machine learning. Front Psychiatry. 2022;13:757961.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRahman A, Mahir SH, Tashrif MTA, Aishi AA, Karim MA, Kundu D, Debnath T, Moududi MAA, Eidmum M. Comparative analysis based on deepseek, chatgpt, and google gemini: Features, techniques, performance, future prospects. arXiv preprint arXiv:250304783 (2025).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"psychological anxiety, large language model (LLM), dynamic interaction, risk assessment","lastPublishedDoi":"10.21203/rs.3.rs-8192788/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8192788/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eIn recent years, psychological anxiety has emerged as a pervasive mental health issue impacting socioeconomic development and individual well-being, making scientific assessment of anxiety cru- cial. While traditional manual evaluation methods and machine learning-based automated assessment approaches each possess unique advantages, they often struggle to simultaneously address efficiency, reliability, and feature generalization capabilities, lacking systematic comparative validation with standard clinical tools. To tackle this challenge, this study first proposes an anxiety risk assess- ment model based on interactive large language models (LLMs). The model is fine-tuned using multi-round psychological counseling dialogue datasets to provide high-quality psychological support.\u003c/p\u003e\u003cp\u003eSubsequently, we introduce a dual-channel BERT framework for anxiety risk assessment, compris- ing a main channel and a high-risk channel. The framework strictly maps symptom dimensions and score ranges from the GAD-7 and HAMA clinical scales. This design enables accurate anxiety risk assessment while promoting psychological intervention through interactive dialogue. Experimental results demonstrate that the proposed model achieves high accuracy (88.86%) and robustness in gen- eral case evaluations, maintains 88.20% accuracy in cross-dataset validation, and exhibits superior performance in identifying high-risk cases (99.34% high-risk recall rate), significantly outperforming advanced baseline models like MentalBERT.\u003c/p\u003e","manuscriptTitle":"Psychological Anxiety Risk Analysis Model Based on Large Language Model Interaction","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-05 01:57:02","doi":"10.21203/rs.3.rs-8192788/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"486411f5-a3e5-4e2d-9567-e6d54cb08e3b","owner":[],"postedDate":"December 5th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-12-05T12:54:10+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-05 01:57:02","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8192788","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8192788","identity":"rs-8192788","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}