Twitter-STMHD: An Extensive User-Level Database of Multiple Mental Health Disorders

Suhavi, Asmit Kumar Singh, Udit Arora, Somyadeep Shrivastava, Aryaveer Singh, Rajiv Ratn Shah, Ponnurangam KumaraguruICWSM 2022

Abstract

This paper introduces Twitter-STMHD, a high-precision, longitudinal dataset comprising over 25,000 Twitter users across eight mental health disorder categories. By providing a user-level view that captures behavior before, during, and after diagnosis, the dataset addresses the limitations of existing "bag-of-tweets" resources. The inclusion of data spanning the COVID-19 pandemic allows for unique insights into how global stressors interact with pre-existing conditions. Preliminary analyses reveal distinct linguistic markers, such as increased self-focus through pronoun usage, and behavioral shifts in topical interests compared to control groups.

Introduction

Social media has emerged as a powerful tool for "digital phenotyping," offering traces of human behavior that can inform mental health research. However, the field has been limited by datasets that often lack temporal context, are restricted to single disorders, or suffer from poor data quality. Clinical psychology relies heavily on patient history; similarly, social media research requires knowing when a diagnosis occurred to distinguish between symptom onset and progression. Furthermore, the impact of global events like the COVID-19 pandemic on specific mental health populations remains under-studied due to a lack of temporally aligned data.

Methods

The Twitter - Self-Reported Temporally-Contextual Mental Health Diagnosis (Twitter-STMHD) dataset was constructed using a multi-stage pipeline designed for high precision:

  1. Anchor Tweet Identification: Users were identified via "anchor tweets" containing explicit self-disclosures of diagnosis (e.g., "I was just diagnosed with ADHD").
  2. Hybrid Annotation: 60% of candidate tweets (approx. 76,000) were hand-annotated and validated by a licensed clinical psychologist, achieving 99.2% precision.
  3. Heuristic Growth: A high-precision lexicon and pattern-matching system, derived from the annotated data, were used to identify the remaining 40% of the dataset automatically.
  4. Temporal Windowing: For each user, a four-year timeline was collected, spanning two years before and two years after the anchor tweet.
  5. Filtering: Strict criteria were applied to ensure data quality, including a minimum of 50 tweets per user and a maximum of 5,000 followers to exclude non-individual accounts.

The Twitter-STMHD Dataset

The primary contribution is the dataset itself, which includes 25,860 unique users across eight categories: ADHD, Anxiety, Bipolar, Depression, MDD (Major Depressive Disorder), OCD, PPD (Post-partum Depression), and PTSD. It also features a control group of approximately 8,000 users. The data is organized into temporal slices representing onset, diagnosis, progression, and a specific COVID-19 period (March 2020 – May 2021).

Linguistic and Behavioral Analysis

The user-level longitudinal nature of the dataset allows for the modeling of a disorder's "journey." Linguistic analysis using LIWC (Linguistic Inquiry and Word Count) revealed significant differences in pronoun usage. Users with mental health disorders, particularly those in the PTSD and Depression groups, utilized first-person singular pronouns ("I", "me") at a much higher rate than the control group, indicating increased self-focus and psychological distress.

Figure 5: Density of pronoun usage across different mental health categories and control users.

Beyond linguistics, the dataset highlights distinct behavioral trends. While control users frequently tweet about "Leisure" activities, users within the disorder categories show a narrowed focus, tweeting significantly more about "Home" and "Money." This suggests a shift away from social and recreational engagement toward more immediate personal or financial concerns.

Figure 6: Density of topical interests (Leisure, Home, Money) across mental health categories and control users.

Additionally, the data confirmed a "night owl" posting pattern among depressed users, where spikes in late-night activity serve as a digital proxy for insomnia. The dataset also captures a massive surge in posting frequency across all groups following the onset of the COVID-19 pandemic in March 2020.

Discussion and Limitations

While Twitter-STMHD provides a massive scale and high precision (94% for automated matching), it is subject to self-reporting bias, as it only includes individuals comfortable with public disclosure. Privacy is managed through a Data Usage Agreement (DUA) and the removal of personally identifiable information. Future work will focus on developing predictive models that leverage this temporal context for early intervention, moving from retrospective observation to proactive mental health support.

Summary updated

BibTeX
@inproceedings{suhavi2022twitter,
  author = {Suhavi and Asmit Kumar Singh and Udit Arora and Somyadeep Shrivastava and Aryaveer Singh and Rajiv Ratn Shah and Ponnurangam Kumaraguru},
  title = {{Twitter-STMHD: An Extensive User-Level Database of Multiple Mental Health Disorders}},
  booktitle = {ICWSM},
  year = {2022},
}
Figure