DP-CARE: a differentially private classifier for mental health analysis in social media posts

Dimitris Karpontinis, Efstathia Soufleri

Keywords: Differential Privacy, Mental Health NLP, Privacy-Preserving Machine Learning, MentalBERT, Stress Detection, AI for Health, Social Media Analysis, Natural Language Processing

Domains: Machine Learning Theory, Language and Learning, AI for Health

TL;DR: We introduce DP-CARE, a lightweight differentially private framework for mental health text classification that preserves user privacy while maintaining strong stress detection performance.

External Link: https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2025.1709671/full

Abstract: We present DP-CARE, a privacy-preserving framework for mental health text classification that combines a frozen domain-specific language model with differentially private optimization. Mental health NLP systems are increasingly used to analyze sensitive user-generated social media content, raising important concerns regarding memorization and unintended disclosure of personal information. DP-CARE addresses these risks by attaching a lightweight classifier to a frozen MentalBERT encoder and training only the classifier using Differentially Private AdamW (DP-AdamW). This design significantly reduces computational cost while providing formal differential privacy guarantees. We evaluate DP-CARE on the Dreaddit stress detection benchmark, where it achieves competitive performance with an F1 score of 78.08% and a recall of 88.67% under a privacy budget of $\epsilon \approx 3$. Our results demonstrate that strong predictive performance can be retained while substantially reducing privacy risks associated with training on sensitive mental health data. DP-CARE provides a practical and deployable approach for privacy-sensitive AI applications in mental healthcare and social media analysis.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 190