Detecting Suicidal Ideation in the Online Environment

Abstract

This thesis explores the detection of suicidal ideation in online environments with a specific focus on bereaved Reddit users. It addresses the overarching research question: How can our understanding of human grief be employed to train Machine Learning (ML) models to automatically detect suicidal ideation among bereaved users in the online environment?

It integrates three perspectives: i) the humanities and social sciences, analyzing how bereaved users with suicidal ideation describe their subjective experiences and identifies the key themes that emerge from these self-expressions; ii) digital technology, investigating the feasibility of applying ML models that integrate Natural Language Processing (NLP) techniques with human interpretations of grief to detect suicidal ideation in bereaved users, and iii) interdisciplinarity, intersecting both the digital and humanities, focusing on ethical considerations associated with employing ML-driven models to data that capture complex human experiences.

This thesis recommends strategies to ensure ethical and reliable practices where ML-driven models are used to detect suicidal ideation: first, to respect users' autonomy and privacy; and second, to explore the possibility of establishing a comprehensive framework that involves specialized advisory boards of AI pioneers, mental health professionals, ethics specialists, policymakers, legal experts, and data protection authorities to facilitate flexible and iterative ethics reviews that enable rapid adaptation to emergent risks posed by evolving ML tools. Such boards should also consider the implications of data sources that exemplify the "privacy paradox", typical of social media; of conducting ongoing research to improve model interpretability; of fostering interdisciplinary collaboration among stakeholders to address potential issues that may arise during both the development and deployment phases; and of maintaining continuous dialogues to clarify accountability and liability for potential misclassifications. These measures aim to balance the benefits of early detection and intervention with the need to protect users' rights and well-being.

This thesis focuses on bereaved, English-speaking Reddit users, but the methodology can be generalized to other cultural contexts or languages. It offers preliminary insights into the online expression of suicidal ideation by bereaved individuals and lays a foundation for future research aimed at refining and ethically implementing automated detection systems for vulnerable populations. By blending humanities-based analysis, ML-driven techniques, and a strong ethical framework, this thesis deepens our understanding of how digital technologies can assist humans in identifying suicidal ideation and contribute to more proactive mental health support. Show more