Scoping Review of Multimodal AI for Suicide Detection on Social Media: Implications for Improvement through Data Justice Lens

Background

Suicide remains a major public health concern among adolescents and young adults, and social media is increasingly a space where signals of distress may emerge before engagement with formal services. Most suicide detection models rely on text-based approaches, overlooking multimodal expressions such as images, emojis, posting rhythms, and behavioral patterns that may be especially salient for youth whose digital communication often relies on visual symbolism and interactional cues. Multimodal AI may offer a more comprehensive approach, yet it remains methodologically fragmented and ethically underexamined. This scoping review synthesizes research on multimodal AI for detecting suicidality on social media, focusing on methodological patterns and ethical implications. Using the Data Justice framework (Taylor, 2017), we discuss how this literature engages questions of visibility, engagement, and non-discrimination.

Methods

Following Arksey and O’Malley’s and Levac et al.’s frameworks and PRISMA-ScR guidelines, we searched six databases for peer-reviewed, English-language studies published between 2010 and 2025. Of 819 records identified, 315 duplicates were removed, and 504 underwent title and abstract screening; 133 were retained for further review. After full-text screening of 16 articles, 12 studies met final inclusion criteria. Data extraction included multimodal configurations, AI methods, prediction targets, labeling and validation procedures, and ethical considerations. Narrative synthesis was conducted due to methodological heterogeneity.

Results

Screening-level analysis (n = 504) confirmed the field’s concentration in text-heavy (88.8%), supervised (90.4%), unimodal (81.8%) paradigms. Among the 12 included studies, text was the primary analytic modality in 11 studies. Visual, behavioral, and metadata features were typically incorporated as secondary inputs and yielded performance gains of 3–10% when combined with linguistic inputs.

Validation was largely limited to internal cross-validation, with little external or temporal testing. Labeling strategies varied substantially, including clinical assessments expert annotation, keyword heuristics, and distant supervision. Ethical safeguards were inconsistently addressed. Most studies relied on publicly scraped data without explicit user consent, raising concerns about contextual privacy and secondary use of sensitive content, and few conducted subgroup fairness analyses.

Collectively, these patterns suggest that current systems are optimized to detect explicit, text-based distress and may overlook indirect, multimodal, or culturally specific expressions of suicidality, particularly among minority populations.

Through the Data Justice lens, three structural gaps emerged: limited visibility due to overrepresentation of English-speaking, text-oriented, and platform-active users with limited geographic and linguistic diversity; weak engagement reflected in minimal consent, contestability, and interpretability of model outputs; and limited attention to non-discrimination, evidenced by absence of systematic fairness evaluation across demographic, linguistic, and cultural groups.

Conclusions

Multimodal AI shows promise for improving suicidality detection but current systems remain shaped by structural barriers towards visibility, engagement, and non-discrimination. Findings highlight the need for expanded non-English NLP resources, interpretable models with human-in-the-loop oversight, and subgroup performance audits with cross-platform validation. For social work research and practice, this study implies that improving suicide prevention requires not only better prediction models but also equitable systems that can identify hidden, indirect, and underrepresented forms of distress in digital environments.