Enhancing Depression Detection through Advanced Text Analysis: Integrating BERT, Autoencoder, and LSTM Models

Abstract

Depression is a austere medical ailment that upsets numerous people worldwide, causing a persistent decrease in mood and significantly impacting their emotions. The article focuses on utilizing BERT techniques and Autoencoders to detect depression from text data, considering gender differences. The work stresses on feature engineering of text data provided by benchmark dataset DAIC_WOZ. We experiment with BERT embeddings that encodes the meaning of text to derive text features. They are then fused with the help of Autoencoders with other parametric features from PHQ-8 survey responses, absolutist word count and gender information. The study found that incorporating this information significantly enhances the functioning of the model. The proposed method outperformed the baseline models. We emphasize the potential of machine learning for mental health research that considers gender differences. We report 98.6% accuracy demonstrated by our method. We found the mean absolute error (MAE) as 0.19 and root mean squared error (RMSE) as 0.282 which signifies the high performance of our proposed method for binary depression classification.