Depression classification on privacy protected facial features data

Abstract

This thesis presents depression classification on privacy protected facial features data. Fast depression classification to help patients receive proper treatment is a method that can prevent the damage of depression. However, fast and effective depression classification is difficult because medical personnel are adequate and the time to analyze depression is long per patient. Applied artificial intelligence in the medical field can help reduce the workload of medical personnel. It is also difficult because of privacy protection. Therefore, we utilize extracted facial features from facial expressions in clinical interview videos to develop a machine learning model. The model utilizes LSTM, attention mechanism, intermediate fusion, and label smoothing approaches to improve performance. The experiments were conducted on 474 video patients collected at Chulalongkorn University. The data set was divided into 134 depressions and 340 non-depressions. Our model achieves 91.67% accuracy, 91.40% precision, 87.03% recall, and 88.89% F1-score. In addition, our model is analyzed using an integrated gradient to explain the important facial features. The significant facial features related to depressive symptoms are head turning, no specific gaze, slow eye movement, no smiles, frowning, grumbling, and scowling, which express a lack of concentration, social disinterest, and negative feelings.