Detection of Mental Health Condition Using Machine Learning Algorithms

This research conducts a detailed study on using machine learning algorithms in mobile applications to predict and detect mental health disorders from text.The first part of the study consists in the analysis of several classification algorithms (Random Forest, Multinomial Nave Bayes, Long Short-Term Memory, Convolutional Neural Network, BERT, Mental BERT, RoBERTa, Mental RoBERTa), using different processing methods (TF-IDF, tokenization, tokenization and GloVe embeddings) and training techniques (using class weights, data augmentation, and regularization techniques), with the intention of using the best model in the mobile application.The benchmark algorithm achieves an F1-Score of around 60%, while the best model reaches 75% for multi-class classification and 96% for the binary model.Two models were finally selected, a binary one based on a LSTM and a multi-class one using BERT transformers.The second part is the mobile application development, having a simple interface, which can serve as a diary and detect potential mental disorders from text input using the selected algorithms.The combination of these parts proves the hypothesis of detection and prevention of mental disorders, helping people learn about mental health in a quick and simple way through an accessible application.