Enhancing explainability and performance of the depression detection model on social media utilizing feature engineering and LLMs

Depression is a mental disorder that negatively affects many people worldwide. The traditional method to help diagnose depression is through questionnaires, and better diagnoses can be obtained by consulting psychiatrists. Because the methods are time-consuming, researchers search for a way to efficiently and effectively diagnose depression before it is too late. Social media data is then utilized for detecting depression since people tend to express their feelings on social media. Different deep learning methods have been proposed for detecting depression on social media. However, the explainability regarding the classification result is limited to identifying negative emotions from the words in the posts. In this research, we aim to enhance the explainability and performance of the depression detection model by utilizing feature engineering and Large Language Models (LLMs). For feature engineering, we compute emotional status, which consists of twelve features representing emotion not just for each post but also before and after the post for some period of time. We also consider language used in the post by using different word embedding models. Different sequence models are also explored, and an attention mechanism is used to help identify the most important posts. The emotional status of the important posts and the classification result are then input to the LLM to describe and relate the emotional status to the characteristics of depression. Based on our experiment, fine-tuning LLM using mental health-related data did help the LLM to produce better explainability by relating easier-understood emotional status to specific characteristics of depression.