Depression detection through transformers-based emotion recognition in multivariate time series facial data

Abstract

Globally, the prevalence of mental health disorders, particularly depression, has become a pressing issue. Early detection and intervention are vital to mitigate the profound impact of depression on individuals and society. Leveraging transformer models, renowned for their excellence in natural language processing and time series tasks, we explore their application in depression detection using multivariate time series (MTS) data from facial expressions. Transformer models excel in sequential data processing but remain relatively unexplored in facial expression analysis. This study aims to compare transformer models applied to first-order time derivative data with traditional methods. We use the distress analysis interview corpus wizard of oz (DAIC-WOZ) dataset and evaluate models with mean absolute error (MAE) and root mean squared error (RMSE) metrics. Results show that transformer models on first derivatives outperform others with an MAE of 4.42 and RMSE of 5.42. While transformer models on raw data surpass XGBoost in RMSE, they fall short of LSTM+transformer with an MAE of 5.41 and RMSE of 6.02. Preprocessing through differentiation enhances transformer models' ability to capture temporal patterns, promising improved depression detection accuracy.