On September 28, 2026

Longitudinal Mental Health Monitoring via Social Media NLP:Detecting Behavioral Drift Using Temporal Transformer Models

Share 0
Tweet 0
Pin 0

The passive, high-frequency nature of social media discourse presents an unprecedented opportunity for continuous,
unobtrusive mental health monitoring at population scale. Conventional clinical paradigms rely on episodic, self-reported
symptom assessments that are susceptible to recall bias, social desirability effects, and temporal sparsity—fundamentally
limiting their capacity to capture the gradual behavioral drift that characterizes the prodromal and deterioration phases of
affective disorders. This paper introduces LMHM-TTM (Longitudinal Mental Health Monitor via Temporal Transformer
Models), a novel computational framework that models the evolution of linguistic and behavioral signatures across multi-
month social media posting histories to detect statistically meaningful drift in mental health state. The architecture employs a
Hierarchical Temporal Transformer (HTT) that encodes individual posts at the token level, aggregates them into session-level
representations using position-aware attention over variable-length temporal gaps, and produces trajectory-level embeddings
via a Longitudinal State Evolution Network (LSEN). Drift detection is formalized as a change-point estimation problem solved
by a differentiable Temporal Divergence Detector (TDD) operating on rolling distributional statistics of the latent trajectory.
Crucially, the system incorporates a suite of twenty-three linguistically-grounded features spanning lexical, syntactic,
pragmatic, network-behavioral, and temporal dimensions, including circadian disruption indices, linguistic complexity decay
curves, and hedonic tone gradients. Extensive experimentation on four longitudinal social media datasets—CLPsych 2015,
SMHD, UMD Reddit Suicidality, and a newly constructed 18-month Twitter/X corpus (DRIFT-18K) comprising 18,247
users—demonstrates that LMHM-TTM achieves an AUC-ROC of 0.924 for depression trajectory detection, F1 of 0.881 for
anxiety state change identification, and a mean early-warning horizon of 23.4 days prior to clinically significant deterioration
events, surpassing all prior longitudinal and cross-sectional baselines by substantial margins. Ethical safeguards, including
differential privacy, consent-aware data governance, and algorithmic fairness auditing, are integral components of the proposed
system.

Ahead of Print

-------------- ---------------
Previous
Next
Book Title

Direct Your Visitors to a Clear Action

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Nibh nisl condimentum id

 Click Here Now

Powered by CookieAdmin
Support