Detection of Suicidal Thoughts on Social Media using Deep Learning

This study focuses the automatic segregation of suicide-related and non-suicide-related text using deep advanced learning techniques that will support mental health analysis and early detection of potential suicidal behavior. For this purpose, several models including SimpleRNN, LSTM, BiLSTM, and GRU have been used for discovering patterns in textual data. This process further included comprehensive text preprocessing through stopword removal, elimination of punctuation and numbers, tokenization, padding and lemmatization to enhance the quality of data and performance of models. Key measures like accuracy, precision, recall, F1score, and confusion matrices were used to assess the models. With an accuracy of 92.65%, LSTM outperformed all other tested models, establishing its superiority in capturing longterm dependencies in text. Its ability to differentiate between content linked to suicide and non-suicide was further validated by visualizations. The promise of deep learning techniques in real-time risk identification and suicide prevention systems is demonstrated by this study.