# An AI-based intelligent diagnosis system for adolescent mental health based on multitask deep learning

## TL;DR

This paper introduces an AI system that uses deep learning to diagnose depression and anxiety in Chinese adolescents based on their written expressions, aiming to overcome limitations of traditional screening methods.

## Contribution

The novel contribution is a multitask deep learning system for adolescent mental health diagnosis using textual data, optimized for Chinese cultural and linguistic context.

## Key findings

- The AI system achieved high correlation (0.706 for depression, 0.693 for anxiety) and strong AUC scores (0.877 and 0.902) on test data.
- Multitask learning improved performance by 6.2%–7.8% in F1-scores and reduced error by 14.2%–18.4% compared to single-task models.
- Data augmentation and adaptations for somatization significantly increased the system’s sensitivity to severe cases.

## Abstract

Adolescent depression and anxiety are becoming increasingly prevalent in China, with rates reaching 20%–30%, driven largely by intense academic pressure and the cultural tendency toward somatization. Traditional screening tools, such as the Patient Health Questionnaire-9 (PHQ-9) and Generalized Anxiety Disorder-7 (GAD-7), often suffer from subjective bias, recall errors, and underreporting due to social stigma. This study developed an AI-based intelligent diagnosis system (IDS) using multitask deep learning to non-intrusively predict comorbid depression and anxiety severity based on the spontaneous textual expressions of Chinese adolescents.

Textual responses from approximately 1,275 adolescents were collected and labeled with clinician-assessed PHQ-9 and GAD-7 scores. Preprocessing involved jieba segmentation and variational autoencoder (VAE)-based data augmentation to address class imbalance, resulting in an expanded test set of 308 samples. The IDS architecture utilizes a Chinese-optimized BERT encoder with self-attention and dual-feature fusion (combining pooled [CLS] tokens and global pooling) to extract shared representations. These are processed through multitask heads for regression (MSE loss) and classification (weighted cross-entropy). The model was trained using an 8:1:1 split with AdamW optimization, cosine annealing, and regularization, supported by ablation studies to validate individual components.

On the test set, the IDS achieved Pearson correlation coefficients of 0.706 for PHQ-9 and 0.693 for GAD-7, with AUC values of 0.877 and 0.902, respectively. Binary classification yielded F1-scores of 0.762 (PHQ-9) and 0.863 (GAD-7). Ablation analysis confirmed that the multitask learning framework improved F1-scores by 6.2%–7.8% and reduced MSE by 14.2%–18.4%. Furthermore, adaptations for somatization and data augmentation for severe cases significantly enhanced the system’s sensitivity.

The IDS offers a robust, culturally sensitive, and scalable tool for adolescent mental health screening. By outperforming single-task baselines, it provides a proactive, privacy-preserving alternative to traditional self-reports. Future research will focus on longitudinal validation, multimodal integration, and ethical deployment strategies to maximize the system’s utility in educational and clinical settings.