The dataset used in this study contains textual expression data from approximately 1,275 adolescents, with each person providing seven to nine independent text responses, where data labels include PHQ-9 scores (0–27 points) and GAD-7 scores (0–21 points), all assessed by professional mental health evaluators according to standard scales. The dataset covers adolescent groups of different age ranges, including junior high school students, high school students, and university students, ensuring sample representativeness and diversity, while the data collection process strictly follows ethical guidelines, with all participants signing informed consent forms, and data anonymization to protect privacy (23). Data preprocessing includes text cleaning, Chinese word segmentation, length control, and intelligent data augmentation steps, where text cleaning mainly removes special characters and normalizes punctuation, while preserving punctuation related to emotional expression to maintain the emotional characteristics of the text. Chinese word segmentation uses the jieba tokenizer for word segmentation processing, with a maximum sequence length set to 96 tokens to balance computational efficiency and information retention. To address common class imbalance issues in mental health data, this study employs intelligent data augmentation technology based on a variational autoencoder (VAE) (24), balancing data distribution across various categories by generating high-quality synthetic samples. The dataset adopts a stratified sampling strategy, divided into training set, validation set, and test set in an 8:1:1 ratio, where stratification is based on the joint labels of PHQ and GAD, ensuring consistency of category distribution across subsets. This sampling strategy effectively avoids data distribution bias, ensuring objectivity and reliability of model evaluation. The original test set contained 127 samples. For comprehensive evaluation and to ensure sufficient sample size for robust performance assessment, the test set was expanded to 308 samples through data augmentation, maintaining the same distribution characteristics as the original test set. Data distribution statistics for the original dataset are shown in Table 1.

The multitask deep learning model proposed in this study is built based on the BERT pretrained model. This model adopts an end-to-end training approach, capable of simultaneously learning prediction tasks for depression and anxiety symptoms, using BERT-base-Chinese as the base encoder to extract deep semantic features from text. The BERT architecture is based on the encoder part of Transformer, employing bidirectional self-attention mechanisms to fully understand contextual information in text (25). The BERT model contains 12 layers of Transformer encoders, each with 12 attention heads, with a total parameter count of approximately 110 million. The self-attention mechanism captures long-distance dependencies in text by calculating similarity between query, key, and value, which is crucial for understanding complex emotional expressions in mental health text, as shown in Equation 1.

where X is the input text sequence, H∈ℝL×d is the hidden state output by BERT, L is the sequence length (maximum length set to 96 tokens to balance computational efficiency and information retention for adolescent psychological expressions), and d = 768 is the hidden layer dimension (standard BERT-base configuration that provides sufficient representational capacity for capturing complex emotional patterns in mental health text). BERT’s pretraining tasks include Masked Language Modeling (MLM) and Next Sentence Prediction (NSP), enabling it to learn rich language representations that are particularly effective for understanding contextual emotional expressions in Chinese adolescent mental health text. We add additional self-attention layers based on BERT encoding to enhance the capture of key information. This mechanism automatically learns dependency relationships between different positions in text, which is particularly suitable for capturing emotional expressions and key information in mental health text, where emotional indicators may be distributed across different parts of the text and require long-range dependency modeling. The multihead attention mechanism captures text features from different perspectives by parallel computation of multiple attention heads (26), as shown in Equation 2:

where Q, K, and V are the query, key, and value matrices, respectively, and d_k is the dimension of the key. The multihead attention mechanism allows the model to simultaneously attend to different representation subspaces, enabling comprehensive capture of diverse emotional patterns and contextual information that are crucial for understanding complex psychological states in mental health text, as shown in Equation 3:

where head_i = Attention (QWiQ,KWiK,VWiV), and h=12 is the number of attention heads. We combine BERT’s pooled output (representing the [CLS] token embedding that captures global semantic information) and global pooling features (obtained by averaging all token embeddings to capture overall text characteristics) to create a comprehensive feature representation. This dual-feature fusion strategy captures both sentence-level semantics and token-level patterns, which is particularly important for mental health text where both global emotional tone and specific emotional keywords contribute to assessment. Feature enhancement is then performed through a multilayer perceptron, which adopts multilayer non-linear transformations capable of learning complex feature combination patterns that are specific to mental health assessment tasks, as shown in Equations 4–5:

where MLP contains two fully connected layers, with an intermediate layer dimension of 512 and an output dimension of 256. The model contains four output heads for regression and classification tasks for both PHQ-9 and GAD-7, respectively. This dual-task, dual-output design enables comprehensive assessment (Equations 6–9): regression tasks directly output continuous scores for fine-grained severity evaluation, while classification tasks divide scores into different severity levels (mild, moderate, severe) for categorical risk assessment. This design simultaneously handles regression and classification tasks, fully utilizing task correlations and the complementary nature of continuous and categorical assessments to improve overall model performance and clinical utility.

where classification tasks adopt a three-class strategy: mild (0–4 points), moderate (5–9 points), and severe (10 points and above).

To address common class imbalance issues in mental health data, this study developed intelligent data augmentation technology based on the VAE. This technology balances data distribution across various categories by generating high-quality synthetic samples while maintaining semantic consistency and emotional expression characteristics of the text. The text VAE model adopts an encoder–decoder architecture, where the encoder maps input text to a probabilistic latent space and the decoder reconstructs text from latent representations. The encoder employs a bidirectional LSTM structure, which is particularly effective for capturing contextual information in Chinese text where word order and context significantly influence meaning. VAE learns latent representations of text through variational inference, which enables the model to generate diverse text samples while maintaining semantic consistency by learning a smooth and continuous latent space that captures the underlying distribution of mental health text patterns, as shown in Equations 10–11.

where q(z|x) is the encoder that maps input text x to a latent space distribution, p(x|z) is the decoder that reconstructs text from latent representations, and μ(x) and σ2(x) are the mean and variance of the latent distribution, respectively. The latent space dimension is set to 128, achieving a balance between information retention (sufficient capacity to encode semantic information) and generation quality (manageable dimensionality for stable training). This dimension choice is based on empirical analysis showing that 128 dimensions can effectively capture semantic features of mental health text while avoiding overfitting and maintaining generation diversity. The VAE model combines mental health label information for text generation, ensuring consistency between generated text and original labels through conditional generation, as shown in Equation 12.

where y is the mental health label (PHQ-9 and GAD-7 scores). This conditional generation formulation enables the VAE to learn the mapping relationship between labels and text during training, where the model learns to associate specific emotional states (encoded in PHQ-9 and GAD-7 scores) with corresponding linguistic patterns. During generation, label guidance ensures that the emotional state of generated text is consistent with target labels, allowing controlled generation of text samples with specific severity levels for addressing class imbalance issues in mental health datasets. Generated text quality is ensured through four mechanisms: 1) pretrained semantic similarity models calculate similarity between generated and original text, with threshold set above 0.7 to ensure semantic consistency; 2) trained classifiers verify that predicted labels of generated text are consistent with original labels, ensuring generation quality and label preservation; 3) random sampling of different regions in latent space, combined with temperature parameters, regulates diversity of generated text while maintaining semantic coherence; and 4) language models evaluate grammatical correctness and fluency of generated text, ensuring natural language quality. The effectiveness of VAE data augmentation technology is evaluated through the metrics in Table 2.

The core of the proposed architecture is a joint mapping function ℱ designed to exploit the clinical correlation between depression (D) and anxiety (A). Formally, we define the parameter space as Θ={θsh,θD,θA}, where θsh represents the shared parameters in the BERT encoder and dual-feature fusion layer, while θ_D and θ_A denote task-specific parameters for depression and anxiety, respectively. The enhanced latent representation serves as a shared semantic bridge and can be expressed as Equation 13:

This shared representation facilitates “inductive transfer,” where features learned for one task provide supplementary predictive signals for the other. Based on this shared representation, the task-specific predictions for depression and anxiety are defined as follows in Equation 14:

The model is jointly optimized by minimizing the weighted sum of task-specific losses, which is formulated as Equation 15:

This joint optimization mechanism acts as a latent regularizer, effectively reducing the hypothesis space and mitigating overfitting—a critical advantage when modeling nuanced psychological text.

We employ a weighted loss function to balance regression and classification tasks. The total loss function integrates regression loss, classification loss, and regularization terms, and is defined as Equation 16:

where α, β, and γ are the weight coefficients for each loss term, determined through grid search optimization. The regression loss is computed as the mean squared error over both PHQ-9 and GAD-7 predictions, as given in Equation 17:

To address class imbalance, the classification loss is formulated using a weighted cross-entropy function, as described in Equation 18:

Where wc is the class weight, calculated based on the sample count of each class in the training set: wc=NNc·C. An L2 regularization term is introduced to prevent overfitting, which is expressed as Equation 19:

where λ = 0.01 is the regularization coefficient, and Θ is the model parameter set. The influence of this weighted loss formulation on the multitask deep learning architecture is crucial for clinical utility. In this study, we empirically set β = 6.0 and α = 0.3 (β ≫ α), creating a mathematical bias toward high-stakes risk categorization. This design ensures that the shared representation F_enhanced is optimized to prioritize features that distinguish between clinical severity levels (e.g., “mild” vs. “severe”). From a psychiatric perspective, identifying categorical risk levels is of higher priority for early-warning systems than achieving marginal gains in precise score regression. Furthermore, the integration of L_{regularization} (γ = 0.01) stabilizes the multitask gradients, ensuring that neither task dominates the parameter updates during backpropagation, thereby enhancing the overall robustness of the multitask framework.

We implement various advanced training strategies to optimize model performance, ensuring the model achieves optimal results in mental health assessment tasks. The AdamW optimizer is combined with a cosine annealing learning rate scheduling strategy, with the learning rate updated according to Equation 20:

where lr_max = 1 × 10⁻⁵ (initial learning rate), lr_min = 1 × 10⁻⁶ (minimum learning rate), T = 35 is the total training epochs, and t is the current epoch. This cosine annealing strategy gradually reduces the learning rate from maximum to minimum over the training process, enabling fine-grained parameter tuning in later epochs while maintaining training stability in early epochs, which is particularly important for fine-tuning pretrained BERT models on mental health text data. Multiple regularization techniques are applied to prevent overfitting: 1) dropout is applied in MLP layers and attention layers with probability set to 0.5, effectively preventing overfitting; 2) weight decay employs L2 regularization with coefficient set to 0.03, preventing parameters from becoming too large; 3) stochastic depth technology randomly skips certain layers in deep networks with probability set to 0.1, improving training stability; 4) label smoothing is applied in classification tasks with smoothing factor set to 0.1, improving model generalization ability; and 5) gradient noise injection is adopted with noise scale of 1 × 10⁻⁵, improving model robustness. This comprehensive regularization strategy creates a synergistic effect that prevents overfitting while maintaining model performance. An early stopping mechanism is implemented based on the validation set F1 average score with patience = 3. The model with the best validation set performance is saved as the final model. Three model ensemble strategies are employed to improve prediction performance, as summarized in Table 3.

Equation 21 shows that the EMA model employs an exponential moving average to update parameters, creating a smoothed version of model weights that reduces training noise and improves generalization:

where β = 0.999 is the smoothing coefficient. This high smoothing coefficient (β = 0.999) ensures that the EMA model weights change gradually, incorporating only a small fraction (1 − β = 0.001) of the current model weights at each update, which creates a more stable and robust model representation that is less sensitive to training fluctuations and better suited for clinical applications requiring consistent predictions.

To analyze the contribution of individual components in the proposed architecture and compare its performance against standard deep learning configurations, we designed a series of experiments. Starting from the full proposed model, several variant models were constructed by selectively removing specific components while keeping all other settings unchanged. These variants serve as benchmarks for standard methods: the removal of the multitask module corresponds to standard single-task learning; the replacement of the enhanced attention reflects a vanilla Transformer-based approach; and the exclusion of VAE represents models trained on unaugmented, imbalanced clinical data. All models were evaluated under identical experimental settings to ensure a fair comparison within the same environment.

The IDS was evaluated on a held-out test set of 308 samples from Chinese adolescents across junior high, high school, and university levels. Performance was assessed across regression (continuous PHQ-9 and GAD-7 total score prediction) and classification (binary clinical screening and PHQ-9 severity stratification) tasks for comorbid depression and anxiety. All experiments were conducted on an NVIDIA RTX 4090 GPU using PyTorch 2.0.1, with hyperparameters optimized via grid search on the validation set (Table 4).

The IDS achieved robust regression performance on clinician-assessed continuous scores (Table 5). For PHQ-9 prediction, the model yielded an MSE of 117.60, RMSE of 10.84, MAE of 8.92, Pearson correlation coefficient (r) of 0.706, and R² of 0.498. Comparable performance was observed for GAD-7 (MSE: 91.58, RMSE: 9.57, MAE: 7.34, r: 0.693, R²: 0.480). AUC values of 0.877 (PHQ-9) and 0.902 (GAD-7) indicated strong discrimination of clinically relevant symptom thresholds.

As shown in Figure 1, predicted scores closely aligned with true values along the identity line for both tasks, with minimal dispersion. This tight clustering confirms the model’s high fidelity in capturing continuous symptom severity from spontaneous text, particularly in the context of culturally mediated somatization patterns.

Binary classification performance is summarized in Table 6. The model attained an accuracy of 0.773 and macro F1-score of 0.762 for PHQ-9, with superior results for GAD-7 (accuracy: 0.838, macro F1-score: 0.863).

Figure 2 provides a normalized multidimensional overview integrating classification (accuracy, F1-score) and regression (inverse MSE) metrics. The substantially larger enclosed area for GAD-7 demonstrates its superior performance across nearly all evaluated dimensions compared to PHQ-9. This holistic advantage further supports the model’s enhanced capability to detect anxiety-related linguistic markers, particularly somatic expressions prevalent among Chinese adolescents.

The confusion matrices (Equation 22) further elucidate task-specific discrimination. For GAD-7, the sensitivity is 0.845 and the specificity is 0.826. In contrast, PHQ-9 exhibits more balanced but lower sensitivity and specificity values.

Severity-stratified analysis for PHQ-9 (Table 7) revealed consistent overall performance, with the highest F1-score in mild cases (0.815) and the lowest in moderate cases (0.698), attributable to greater symptom heterogeneity in the moderate range. High specificity across all strata supports effective rule-out capability for screening applications.

Training proceeded for 35 epochs with early stopping (patience = 3). As depicted in Figure 3, total loss, PHQ-9 MSE, and GAD-7 MSE exhibited monotonic decline, with negligible train-validation gaps after epoch 10. The accompanying learning rate schedule ensured stable optimization. Asymmetric loss weighting prioritized classification while maintaining regression accuracy, confirming robust multitask convergence without overfitting.

Ablation results (Table 8) confirmed the contribution of each component. The absence of multitask learning substantially degraded F1-scores (6.2%–7.8%) and increased MSE (14.2%–18.4%). Removing culturally attuned somatization adaptation reduced GAD-7 F1-score by 5.1%, while omitting severe-case augmentation lowered recall in high-risk subgroups by 4.4%–6.3%.