Digital Phenotypes of Suicide Risk on Reddit: Interpretable NLP Signals of Social Disconnection and Help-Seeking

Background

Suicide risk frequently emerges in digital environments before contact with formal care systems, positioning online platforms as critical sites for early identification and intervention. While prior natural language processing (NLP) studies demonstrate strong predictive performance in detecting suicide-related content, they largely prioritize classification accuracy over interpretability and rarely anchor linguistic patterns in social theory. This limits their relevance for social work practice, where understanding mechanisms of distress and engagement is essential. Drawing on Digital Phenotyping and the Interpersonal Theory of Suicide (specifically the constructs of thwarted belongingness and perceived burdensomeness; Joiner, 2005), this study examines whether suicide-related language reflects a theoretically grounded and interpretable digital phenotype characterized by social disconnection and perceived burdensomeness.

Methods

We analyzed a publicly available Reddit dataset (N = 15,477 English-language posts) labeled as suicidal (n = 7,275) or non-suicidal (n = 8,202). Posts were constructed by combining titles and text bodies. We developed interpretable linguistic indicators representing digital phenotypes, including post length, first-person pronoun use, negation, negative lexical density, belongingness-related language, burdensomeness-related language, help-seeking expressions, and relational references. Group differences were assessed using Mann–Whitney U tests and Cohen’s d effect sizes. Predictive analyses compared: (1) TF-IDF + Logistic Regression, (2) TF-IDF + Linear SVM, (3) Interpretable feature-based Logistic Regression, Model performance was evaluated using accuracy, precision, recall, specificity, F1-score, ROC-AUC, and PR-AUC.

Results

Suicidal posts were significantly longer (M = 165.5 vs. 126.2 words; d = 0.23, p < .001), indicating more elaborated expressions of distress. They showed substantially higher levels of negative lexical content (d = 0.78), negation (d = 0.67), and first-person language (d = 0.46), reflecting intensified negative affect and self-focus. Importantly, suicidal posts demonstrated elevated belongingness-related language (d = 0.26) and burdensomeness-related language (d = 0.33), providing empirical support for core constructs of the Interpersonal Theory of Suicide. Notably, help-seeking expressions were also significantly higher (d = 0.22), suggesting that online suicidal discourse includes active attempts to engage support rather than solely passive distress expression. In predictive analyses, TF-IDF + Linear SVM achieved the highest performance (F1 = .938, ROC-AUC = .987), followed by TF-IDF + Logistic Regression. The interpretable feature-based model showed lower but meaningful performance (F1 = .806, ROC-AUC = .884), indicating that transparent, theory-informed linguistic features capture substantial components of suicide-related expression.

Conclusions and Implications

Suicide-related language on Reddit reflects a coherent and theoretically grounded digital phenotype characterized by negative affect, social disconnection, perceived burdensomeness, and active help-seeking. These findings extend prior NLP research by demonstrating that interpretable, theory-aligned features can meaningfully represent suicide risk signals in digital environments. For social work, this approach supports earlier identification of unmet needs, enhances interpretability for ethical decision-making, and provides a pathway to integrate digital data into prevention strategies targeting individuals who remain outside formal care systems.