Author Name: Prakhar Shankar Date: 27-03-2026
Social media platforms have become repositories of psychologically rich behavioral data whose analysis offers unprecedented opportunities for population-level depression screening and individual-level clinical assessment. This paper provides a comprehensive review of passive sensing and natural language processing (NLP) approaches to depression detection from social media data, evaluating their validity, clinical utility, ethical boundaries, and equity implications. The paper reviews the landmark CLPsych shared task series and the foundational studies demonstrating that Twitter, Facebook, Instagram, and Reddit language patterns predict clinical depression diagnoses with AUC values of 0.70–0.92. Key linguistic markers — reduced social references, increased self-focused language, elevated negative affect vocabulary, reduced future temporal orientation — are reviewed alongside behavioral markers including reduced posting frequency, decreased network engagement, altered circadian activity patterns, and shift toward more passive consumption. The paper critically evaluates the construct validity of NLP-based depression detection: do these models measure depression or correlated social media behavior patterns? Differential validity analysis reveals concerning disparities — model performance is consistently lower for Black users, non-English speakers, men, and older adults, reflecting training data biases. The paper provides a detailed ethical framework for social media mental health assessment addressing four core tensions: the therapeutic potential versus surveillance risk, opt-in consent versus population monitoring, accuracy versus equity, and clinical utility versus commercial exploitation. A governance framework for responsible deployment is proposed, including independent algorithmic auditing requirements, equity performance standards, meaningful consent architectures, and therapeutic escalation protocols.
Keywords: depression screening; natural language processing; passive sensing; social media mental health; algorithmic fairness; ethical AI; digital phenotyping; NLP mental health.
