LAUSR

Patients with severe depressive symptoms are exposed to high suicidal risk. Since assessing depressive symptoms is either implicated by variability among subjects and clinicians or by time-consuming procedures, it is difficult to do objectively and effortlessly. There is a need for automatic approaches to estimate them. Previous investigations on speech-based severity’s categorical assignment predominantly focused on distinguishing depressive disorders from healthy. This article presents an automatic assessment system of severe self-reported depressive symptoms (SSDSs) to classify speakers with severe depressive symptoms from normal and not so severe ones. First, modulation-domain spectral centroid mean (MSCM) features characterizing severe depressive symptoms are extracted. Subsequently, aiming at the problems of limited and imbalanced training data, we establish variance-weighted sum-to- ${H}$ constraint ( ${H}$ is an adjustable parameter) collaborative representation-based classification (VWSC-CRC) method to exploit interclass separability between SSDS and non-SSDS. The proposed system (MSCM–VWSC-CRC) was evaluated on the small and imbalanced data set from AVEC2013 by dominant metrics, such as F1-score and area under the precision–recall curve (AU-PRC), as well as auxiliary indices, including specificity, recall (sensitivity), precision, accuracy, and area under the receiver operating characteristic curve (AU-ROC) as needed. The results exhibit a clear advantage over three important systems in the literature. The gains of this article are likely to lay the substantive groundwork to assist clinicians in automatically screening subjects with severe depressive symptoms so as to facilitate the diagnosis of the corresponding psychological disorders.