Leveraging AI for Early Detection of Depression from Text

Abstract

Depression is a highly common mental disorder. Depression is characterized by depressed mood or loss of interest in activities for a duration of time. Despite being so prevalent, early detection is often hampered by stigma and unawareness. This project is to develop an application that analyzes users' written responses to mental health questions to detect depression and provide real-time predictions. The suggested methodology is collection of user inputs in the form of responses to predefined questions regarding mental health through a mobile application. The text inputs are fed into NLP-based tokenization, lemmatization, and vectorization (TF-IDF and Word2Vec) processes. A hybrid approach is taken for ensemble learning, using a traditional machine learning model (LinearSVC) and a deep learning model (BiLSTM), to process the input and provide real-time depression predictions through hard voting. Experimental results show that the ensemble model is better than individual models in F1-score and accuracy, at a maximum accuracy of 93% in classifying different types of multiple depression conditions like Major Depressive, Bipolar, Postpartum, and Atypical Depression. The system is interfaced directly in a Flutter-enabled mobile app so that the users can self-test and safely store results for regular mental health checkups. The project makes it possible for individuals to continuously monitor their mental wellbeing, facilitate early intervention, and facilitate contact with medical professionals