Early detection of mental health on social media using a hybrid Bi-LSTM-XGBoost model: a comparative study

The case of mental health disorders has been a main topic in the clinical and psychological field. The advancement of computing studies, especially in Natural Language Processing (NLP)-a subset of Machine Learning, created a system of detection that can detect the mental health state of a person in early stage to prevent the eventuality of the worst case. This is crucial since there has been a lot of case of mental health disorder-such as depression and suicide, remains undetected and untreated-especially when the internet usage is more prevalent than ever even among the most vulnerable users, which are the preadolescent users. This study explores the models that can accurately predict mental health disorder with the provided six labels the model can predict. The labels are anxiety, depression, personality disorder, stress, bipolar, and normal. The dataset is gathered from a Kaggle repository which is then processed and refined further for the training process. From multiple evaluations across diverse amount of texts from different users, our Bi-LSTM-XGBoost model outperforms the other models with an accuracy of 0.9035 and 0.4320 loss, while other models fall short within 50-84% accuracy. Further improvement can be made with our model, whether from improving the model's parameters further or by improving the quantity and quality of the dataset gathered.