Mental Health Symptoms Detection System from Social Media Posts using FastText Embedding and an Ensemble Deep Learning Model

Abstract

Social media is becoming a significant tool for early emotional and psychological identification because of mental health disorders, like depression and anxiety, are increasingly prominent in online social media. This paper presents an ensemble deep learning approach to detect potential indicators of psychotic behavior—such as depression, anxiety, and related mental health conditions—based on user-generated textual data. The suggested method combines FastText word embeddings with a hybrid CNN+LSTM ensemble model to capture local semantic patterns and long-term contextual dependencies. To enhance dataset diversity and linguistic representativeness, data were gathered from Twitter and Reddit which encompass a wide range of user expressions and linguistic variances. A number of experiments were performed with optimized hyperparameters and compared them to conventional deep learning models and a fine-tuned BERT transformer. The suggested model got an overall accuracy of 91.89%, which is quite close to BERT's 93.10% but it maintained significantly lower computational cost and training time. The model was trained and evaluated using MATLAB tool. In order to check the performance of proposed model, evaluation matrices like accuracy, recall, F1-score were used that shows that suggested system is both accurate and morally sound. The findings highlight the potential of lightweight yet robust architectures for large-scale, real-time mental health monitoring. it offers a promising tool for early identification of mental health concerns and contributes to the development of intelligent, non-invasive monitoring systems for online psychological well-being assessment.