Using language in social media posts to study the network dynamics of depression longitudinally

Abstract

Background: Network theory of mental illness posits that causal interactions between symptoms give rise to mental health disorders. This is thought to occur because positive feedback loops between symptoms trigger cascades of further symptom activation. Increasing evidence suggests that depression network connectivity is therefore a risk factor for transitioning and sustaining a depressive state. However, much of the evidence comes from cross-sectional studies that estimate networks across groups, rather than within individuals. We used a novel method to construct personalised depression-relevant networks from social media data to test the hypothesis that network connectivity is linked to depression severity and increases during a depressive episode. Methods: We analysed Twitter data from 946 participants who retrospectively reported the dates of any depressive episodes they experienced in the past 12 months and self-reported current depressive symptom severity. Daily Tweets were subjected to textual analysis, which allowed us to construct personalised, within-subject, depression networks, based on 9 a priori text features previously associated with depression severity. We tested for associations between network connectivity and current depression severity and, in participants who experienced a depressive episode in the past year, we tested if connectivity increased during the dates of a self-reported episode (N = 286). Results: Significant bivariate associations were found between current depression severity and 8/9 of the text features examined. In line with our hypothesis, individuals with greater depression severity had a significantly higher overall network connectivity between these features than those with lesser severity (β = 0.008, SE = 0.003, p = 0.002). Importantly, we observed within-subject changes in overall network connectivity associated with the dates of a self-reported depressive episode (β = 0.03, SE = 0.009, p = 0.005). Conclusions: The connectivity within personalized depression networks changes dynamically with changes in current depression symptoms. Social media data provides a fruitful, albeit noisy, source of data to test key within-subject predictions of network theory.

Machine learning of language use on Twitter reveals weak and non-specific predictions

“…This necessitates taking a ‘trait’ perspective on the mental health symptoms we assessed and it is likely that our model is diluted by variations in state/episodic features of depression. However, in a recent study, we found that individuals’ use of depression-relevant text features in fact didn’t change significantly across within-subject periods of mental health and wellness, suggesting this may not be a major issue 63 .…”

“…Twitter data, by itself, has already proven an interesting testbed for nascent theories of mental health such as network theory, which for example, has struggled to acquire large enough longitudinal datasets to test some of its core predictions 62 . We recently found for example that using social media posts as a proxy for experience sampling allowed us to study a large cohort of individuals through a transition to a depressed state, detecting subtle network signatures of depression vulnerability 63 .…”

Machine learning of language use on Twitter reveals weak and non-specific predictions

“…This necessitates taking a 'trait' perspective on the mental health symptoms we assessed and it is likely that our model is diluted by variations in state/episodic features of depression. However, in a recent study we found that individuals' use of depression-relevant text features in fact didn't change significantly across within-subject periods of mental health and wellness, suggesting this may not be a major issue [72].…”

“…Twitter data, by itself, has already proven an interesting testbed for nascent theories of mental health such as network theory, which for example, has struggled to acquire large enough longitudinal datasets to test some of its core predictions [71]. We recently found for example that using social media posts as a proxy for experience sampling allowed us to study a large cohort of individuals through a transition to a depressed state, detecting subtle network signatures of depression vulnerability [72].…”

Using natural language from a smartphone pregnancy app to identify maternal depression

“…To the best of our knowledge, this is one of the rst prospective longitudinal studies to use natural language collection 50 and the rst focused on maternal depression symptom prediction. Incorporating language inputs enables moderate predictive ability of depressive symptoms among peripartum patients in a large academic health system.…”