A Deep Learning Framework for Early Detection of Depression from Twitter Data Using Topic Modeling and Emoji Sentiment Features

Abstract

The early detection of depression has become more significant as mental health disorders continue to impact individuals world-wide. The popularity of social media platforms presents an opportunity to analyse users' emotional expressions and behavioural patterns for timely depression assessment. In this paper, we propose a hybrid deep learning framework for detecting depression using Twitter data. Our framework integrates Natural Language Processing (NLP), topic modelling, emoji sentiment analysis, and machine learning. First, the tweets are pre-processed using text normalization, tokenization, stop-word removal and lemmatization for improved data quality. Textual feature representations are generated using TF-IDF, hidden semantic topics are detected using Latent Dirichlet Allocation (LDA), and emotional signals are extracted from user posts using emoji sentiment analysis. The extracted features are used to train Logistic Regression, Random Forest, Long Short-Term Memory (LSTM) and Distil BERT models. We measure the performance by using Accuracy, Precision, Recall, F1-score and Receiver operating characteristic-area under the curve (ROC-AUC). The experimental results show that the Distil BERT model achieves the best classification results with an accuracy of 93.2%, which outperforms the other classifiers, due to its ability to efficiently capture contextual information. The proposed framework presents a scalable and reliable solution for automated depression detection from social media data and can potentially assist in early mental health monitoring and clinical decision making.