Review of Advancements in Depression Detection Using Social Media Data

Abstract

A large population embraced social media to share thoughts, emotions, and daily experiences through text, images, audio, or video posts. This user-generated content (UGC) serves various purposes, including user profiling, sentiment analysis, and disease detection or tracking. Notably, researchers recognized the potential of UGC for assessing mental health due to its unobtrusive and real-time monitoring capabilities. Recent reviews on depression identification from textual UGC using AI models covered tools and techniques but overlooked critical components such as datasets, lexicons, features, and subtasks, which are essential for understanding the progress and tasks undertaken. This survey adopts a systematic approach and formulates five research questions to examine the relevant literature concerning these elements. Additionally, it organizes machine learning and deep learning (ML/DL) training features from textual UGC in a hierarchical manner and maps the literature on depression detection into various subtasks. The review highlights that despite the prevalence studies, datasets are limited in both quantity and size, with many relying on less reliable ground truth collection methods such as self-reported diagnosis statements (SRDS). Furthermore, the review identifies an overemphasis on certain textual features, such as n-grams and affective elements, while others, such as life events, egocentric graphs, and intervention/coping style, remain largely unexplored. It is crucial for practical AI depression detection systems to develop expertise in tasks such as severity, symptom detection, and explainable/interpretable depression analysis to instill confidence and trust among users.