Title: FedMental: Assessing Federated Learning for Mental Health Identification via Social Media Content

Abstract: Machine Learning (ML) models frequently leverage text from social media to flag users displaying severe mental health risks. Nevertheless, the dissemination of such sensitive information introduces significant privacy concerns and hinders the expansion of benchmark datasets. This study rigorously examines whether privacy-preserving ML methods can facilitate safer data exchange without compromising model efficacy. We implement federated learning (FL) and Differentially Private FL to address two prominent mental health prediction challenges: identifying depression on X (formerly Twitter) and detecting suicide crises on Reddit. By modeling each user as a distinct client within a non-IID framework, we simulate practical data-sharing environments, testing various aggregation methods, client participation rates, and privacy budgets. Our results indicate that standard FL delivers performance metrics similar to centralized training for depression classification (centralized F1 score of 85.63 versus the top FL model’s 83.16). However, Differentially Private FL exhibits a substantial trade-off between privacy and accuracy, suffering an F1 score decrease of up to 27.01 even under modest noise conditions (epsilon = 50). This performance degradation stems from the corruption of critical but sparse linguistic indicators associated with mental health, such as specific health-related topics and emotional vocabulary. These findings empirically highlight both the capabilities and constraints of existing privacy-preserving approaches in the context of mental health inference.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC