Exploring Emotion Timeline Patterns in Social Media for Automatic Identification of Depression and Anxiety

Over the past years, public awareness of mental health issues has increased, and people are paying more attention to their mental well-being. Depressive and anxiety disorders are among the key causes of the global disease burden. These two mental conditions can disrupt the daily lives of people affected by them and, in severe cases, depression may even lead to suicide. To prevent suicide and alleviate the burden on the affected individuals, researchers have increasingly focused on automatically detecting these mental conditions using social media data, as people with mental health issues often seek supports and share their feelings on social media. Since depression and anxiety are chronic issues and related to emotions, past studies have utilised emotion information to identify social media users with depression and anxiety. However, these emotion features used in the past studies are often static, lacking the temporal fluctuation patterns that can effectively reflect the characteristics of these mental conditions. To address this issue, I propose a novel method that applies the timeline Shapelet classification method, a time series analysis method, to the automatic detection of depression and anxiety based on emotion intensity timeline patterns. This study involves two languages. I first created a new Thai dataset for depression and anxiety detection. Next, I selected and modified an existing English dataset for depression detection. Both datasets contain users' tweeting histories in X. Tweets in the Thai dataset were labelled with 26 emotion intensities by both human annotators and a Large Language Model (LLM). In contrast, tweets in the English dataset were automatically labelled with intensities of four emotions using an LLM. Based on the manually annotated Thai dataset, depression and anxiety are found to be associated with 14 and 10 emotions respectively. My study demonstrates that Shapelet classifiers can identify emotion intensity timeline patterns associated with depression and anxiety. Using these patterns, the classifiers detected depression with a precision of 0.7292 on the English dataset, and achieved maximum precision of 1.00 in detecting both depression and anxiety on the Thai dataset. These classifiers also outperformed benchmark models on the Thai dataset in detecting the two conditions, although they produced relatively low recall on both datasets. Finally, this study also investigates the potential of porting emotion intensity timeline patterns for cross-lingual depression detection, from English to Thai in this case. In my experiment, the Shapelet classifiers trained on English dataset could effectively detect depression in the Thai dataset, showing the potential of this approach for cross-lingual application, particularly for under-resourced languages.