Detecting Depression in Social Media with NLP Models Trained on Journal Entry Data

Abstract

Writing has always been recognized as a powerful means of expressing human emotions, serving as a reflective practice that allows individuals to process and articulate their inner experiences. However, with the rise of social media, the landscape of emotional expression has shifted. This transition from private journaling to public social media posting raises important questions about how effectively these platforms serve as emotional outlets and what they reveal about users' mental health, specifically with pervasive mood disorders like depression, which affects over 18 million adults in the United States. Recently, NLP models have been noted as a promising tool for detecting underlying sentiment in text. This research explores how the Twitter posts of individuals suffering from depression compare when analyzed using a natural language processing (NLP) model trained on journal data classified by emotions. Two separate clustering approaches were used to reduce dimensionality in training data and train machine learning models, one with spectral clustering and principal component analysis (PCA), and the other with the Natural Language Toolkit (NLTK) library. The results of both machine learning approaches, with accuracy over 99%,  demonstrated that tweets of depressed Twitter users are classified as more negative compared to those of non-depressed users. These findings suggest that the emotional content expressed in social media posts by individuals with depression is consistently more negative, aligning with the patterns observed in their journal entries. Ultimately, this research highlights the evolving role of social media as a platform for emotional expression and its implications for mental health monitoring.