Predicting Affective Dimensions Based on Self Assessed Depression Severity

Depression is a state of severe despondency and affects a person’s thoughts and behavior. Depression leads to several psychiatric symptoms such as fatigue, restlessness, insomnia as well as other mood disorders (e.g. anxiety and irritation). These symptoms have a resultant impact on the subject’s emotional expression. In this work, we address the problem of predicting the emotional dimensions of valence, arousal and dominance in subjects suffering from variable levels of depression, as quantified by the Beck Depression Inventory-II (BDI-II) index. We investigate the relationship between depression severity and affect, and propose a novel method for incorporating the BDIII index in affect prediction. We validate our models on two datasets recorded as a part of the AViD (Audio-Visual Depressive language) corpus: Freeform and Northwind. Using the depression severity and a set of audio-visual cues, we obtain an average correlation coefficient of .33/.52 for affective dimension prediction in the Freeform/Northwind datasets, against baseline performances of .24/.48 based on using the audio-visual cues only. Our experiments suggest that the knowledge of depression severity significantly improves the emotion dimension prediction, however the BDI-II score incorporation scheme varies between the two datasets of interest.