Suicide Risk Assessment via Temporal Psycholinguistic Modeling (Student Abstract)

Puneet Mathur, Ramit Sawhney, Rajiv Ratn Shah

Abstract

Automated suicide ideation detection often fails by analyzing social media posts in isolation, leading to high false-positive rates. This study proposes a temporal psycholinguistic model that integrates historical tweeting patterns—specifically modeling gradual "build-up" and episodic "waves" of depression—with social network structures. By incorporating clinical psychological theory into deep learning architectures, our approach significantly improves precision and F1 scores, providing a more robust framework for real-world mental health intervention.

Introduction

Identifying suicidal intent on social media is a high-stakes task where accuracy is paramount. Most current automated systems struggle because they focus solely on the linguistic content of individual tweets. However, suicidality is rarely a spontaneous event; it is typically the result of a long-term psychological trajectory. Furthermore, social media users exist within "homophily networks" where individuals with similar mental states often interact. By ignoring historical context and community interactions, existing models miss critical cues that differentiate passing remarks from genuine cries for help.

Methodology

The researchers proposed a multi-pronged approach combining deep learning with clinical psychological theory:

  1. Linguistic Pipeline: A BiLSTM (Bidirectional Long Short-Term Memory) network with an Attention mechanism extracts latent vector representations from tweets.
  2. Temporal Modeling: This core novelty models suicidal tendency as a sum of two time-varying functions: an exponential function representing the gradual "build-up" of ideation ($\lambda$) and a Fourier Series modeling repetitive, episodic "waves" of depression ($\mu$).
  3. Graph Convolutional Networks (GCN): To capture social context, a heterogeneous graph of users and tweets was constructed. This allows the model to leverage linguistic homophily—the tendency for users with similar mental states to interact.
Temporal modeling of suicidality showing episodic waves and exponential build-up.

Experiments and Results

The study utilized the SNAP-BATNET dataset, containing 34,306 tweets (3,984 labeled as suicidal) with substantial inter-rater agreement ($\kappa = 0.72$).

The combined model (Text + Build-up + Episodic + GCN) outperformed the state-of-the-art baseline across all metrics. The most striking improvement was in Precision, which jumped from 72.20% to 91.98%. This reduction in false alarms is critical for preventing "alert fatigue" among human moderators and mental health professionals.

Performance comparison of the proposed model components against the SNAP-BATNET baseline.

Discussion and Key Insights

  • Psychology-Informed Math: Using clinical observations to justify exponential and sinusoidal functions bridges the gap between data science and clinical psychology.
  • Context Reduces Noise: Temporal features help suppress false positives by providing the historical perspective necessary to filter out suicidal phrases used out of context.
  • The Power of the Graph: The GCN provided a gain in prediction confidence, suggesting that social interactions are a valid signal for mental health assessment.

Limitations and Future Work

The model currently treats suicide risk as a binary classification, whereas risk exists on a complex spectrum. There are also inherent risks of demographic bias and ethical concerns regarding the use of public data for mental health interventions. Future work aims to move beyond binary labels and refine how these models can be used in clinical settings without infringing on user privacy.

Summary updated

Abstract

Social media platforms are increasingly being used for studying psycho-linguistic phenomenon to model expressions of suicidal intent in tweets. Most recent work in suicidal ideation detection doesn't leverage contextual psychological cues. In this work, we hypothesize that the contextual information embedded in the form of historical activities of users and homophily networks formed between like-minded individuals in Twitter can substantially improve existing techniques for automated identification of suicidal tweets. This premise is extensively tested to yield state of the art results as compared to linguistic only models, and the state-of-the-art model.
BibTeX
@inproceedings{mathur2020suicide,
  author = {Puneet Mathur and Ramit Sawhney and Rajiv Ratn Shah},
  title = {{Suicide Risk Assessment via Temporal Psycholinguistic Modeling (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i10.7209},
}
Figure