Social media-based diagnostic paradigm for affective pattern analysis using deep learning

Abstract The physiological phenomenon of Premenstrual Syndrome (PMS) elicits diverse emotional responses in women. For some, these affective responses intensify into conditions such as Premenstrual Dysphoric Disorder (PMDD) and Premenstrual Exacerbations (PME). Automatically identifying textual expressions of these affective states may help flag content suggestive of underlying mental health conditions, such as Unipolar Depression, Bipolar Disorder, Schizophrenia, and Suicidal Behaviour, for further clinical evaluation. While traditional diagnostic mechanisms have their merits, they can be influenced by bias due to convenience sampling and may not always consider the nuanced, firsthand accounts of these emotional states. With the ascent of digital communication, social media platforms, especially Reddit and YouTube, have emerged as pivotal arenas where women candidly discuss and share their experiences. In our study, we classified posts indicating the presence or absence of symptoms linked to mental illness in PMS with an accuracy of 95%. Additionally, we were able to categorize posts indicating mood disorders into their respective disorder subtypes with an accuracy of 80%. These figures are obtained on a single held-out split of a modest, predominantly lexicon-labelled dataset and should be regarded as indicative. We contend that our findings are a step towards utilizing user-generated content for the large-scale analysis of mental health signals related to PMS. We emphasize that our approach performs automated classification of self-reported affective content and detection of risk signals from text. The resulting labels are content-level indications that may warrant further clinical evaluation. They do not, by themselves, constitute a clinical diagnosis, particularly for conditions such as bipolar disorder and schizophrenia, which cannot be established from isolated textual expressions.