Mental Health Prediction Model on Social Media Data Using CNN-BiLSTM

Abstract

Social media has transformed into a global platform for expression and interaction where users can share photos, images, and videos. The rapid development and widespread use of social media afford the opportunity to analyze the construction of social life in societies and communities. As a result of alterations in lifestyle during the COVID-19 pandemic, mental health disorders increased. Mental health is a complex disease involving numerous individual, socioeconomic, and clinical variables. Natural language processing and analysis methods are required to address this complexity. The classification of mental health-related texts, which can serve as early warnings and early diagnoses, is facilitated by analytical and natural language processing techniques. In this investigation, a CNN-BiLSTM model was utilized, which was aided by a FastText-based word weighting method. The utilized data set consists of texts on mental health with labels such as borderline personality disorder (BPD), anxiety, depression, bipolar, mentalillness, schizophrenia, and poison. There are 35000 training records and 6108 test records. The data will undergo a data cleansing procedure, which will include lower text stages, number removal, reading mark removal, and stopword removal. Modeling with CNN-BiLSTM and FastText weighting yielded an F1-Score and accuracy of 85% and 85%, respectively. In comparison to the Bi-LSTM model, the F1-Score and accuracy were both 83%.