Naturalistic Online Language as a Marker of Depression in Midlife and Older Adults: Computational Text Analysis of Bluesky and Reddit Posts

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Background Social media use among older adults continues to grow. Many people use social media to establish meaningful social ties and discuss their mental health. Depression in midlife and older adults is a critical public health concern, yet scalable, sensitive methods for early detection remain limited. Natural language processing offers new opportunities to examine sentiment and mental health through online language in typically understudied populations. Objective This study examined whether sentiment and cognitive features of naturalistic social media language are associated with depression diagnosis and symptom severity among midlife and older adults. Methods We constructed cohorts of adults aged 50 years and older from Reddit (n=688) and Bluesky (n=210). Participants who had ever had depression and never been diagnosed were recruited via a self-report survey on Prolific that collected depression history, age of diagnosis, current symptoms, and social media handles. Public posts were preprocessed and analyzed using a validated rule-based sentiment analysis (Valence Aware Dictionary and Sentiment Reasoner). Sentiment measures were aggregated at the user level and compared by depression status within and across platforms. Results On Bluesky, users with a history of depression exhibited lower Valence Aware Dictionary and Sentiment Reasoner compound sentiment compared to users who had never been diagnosed (ρ=−0.16; P .05). These findings indicate a platform-specific relationship between depression status and online sentiment. Conclusions Negative sentiment expressed on Bluesky was associated with depression history (ever diagnosed vs never diagnosed) among midlife and older adults. On Reddit, current depression severity showed a small association with negative sentiment that did not remain statistically significant after correction for multiple comparisons. Our findings suggest that sentiment-based language features may capture modest differences in online expression related to depression. Future work with larger samples and longitudinal symptom assessment is needed to determine whether sentiment reliably tracks with depression-related symptoms over time.