Abstract
This paper addresses the AVEC 2017 Depression Sub-Challenge, where the objective is to propose methods which can aid automated prediction of depression severity. In this paper, we specifically focus on biomarkers of psychomotor retardation, which are a key trait of depressive episodes, to propose three sets of methods. We propose a novel set of temporal features (which we called ’turbulence features’) and show their effectiveness. We offer a novel methodology to target specific craniofacial movements indicative of psychomotor retardation and hence of depression. Further, we present a novel method for quantifying abnormalities of speech spectra of individuals with depression using Fisher vector encoding of spectral low level descriptors (LLDs). So far, in the AVEC challenge on prediction of patient health questionnaire (PHQ) scores on the Test set, we achieve a root mean square error (RMSE) score of 6.34 and a mean absolute error (MAE) score of 5.30, both of which are better than the best results on the AVEC test set as given in the baseline paper i.e. 6.97 and 5.66, respectively. This suggests that our method is a viable proof of concept and may lead to fully automated objective depression screening protocols.
