Abstract
INTRODUCTION: Mental health NLP models are increasingly used to detect psychological states such as stress and depression from user-generated social media content. Although transformer based models such as MentalBERT achieve strong predictive performance, they are typically trained on sensitive data, raising concerns about memorization and unintended disclosure of personally identifiable information. METHODS: We propose DP-CARE, a simple yet effective privacy-preserving framework that attaches a lightweight classifier to a frozen, domain-specific encoder and trains it using Differentially Private AdamW (DP-AdamW). This approach mitigates privacy risks while maintaining computational efficiency. RESULTS: We evaluate DP-CARE on the Dreaddit dataset for stress detection. Our method achieves competitive performance, with an F1 score of 78.08% and a recall of 88.67%, under a privacy budget of ε ≈ 3. DISCUSSION: The results indicate that lightweight, differentially private fine-tuning offers a practical and ethical approach for deploying NLP systems in privacy-sensitive mental health contexts. DP-CARE demonstrates that strong predictive performance can be retained while significantly reducing privacy risks associated with training on sensitive user data.