There is a growing interest in computational approaches permitting accurate detection of nonverbal signs of depression and related symptoms (i.e., anxiety and distress) that may serve as minimally intrusive means of monitoring illness progression. The aim of the present work was to develop a methodology for detecting such signs and to evaluate its generalizability and clinical specificity for detecting signs of depression and anxiety. Our approach focused on dynamic descriptors of facial expressions, employing motion history image, combined with appearance-based feature extraction algorithms (local binary patterns, histogram of oriented gradients), and visual geometry group features derived using deep learning networks through transfer learning. The relative performance of various alternative feature description and extraction techniques was first evaluated on a novel dataset comprising patients with a clinical diagnosis of depression (n= 20) and healthy volunteers (n= 45). Among various schemes involving depression measures as outcomes, best performance was obtained for continuous assessment of depression severity (as opposed to binary classification of patients and healthy volunteers). Comparable performance was achieved on a benchmark dataset, the audio/visual emotion challenge (AVEC’14). Regarding clinical specificity, results indicated that the proposed methodology was more accurate in detecting visual signs associated with self-reported anxiety symptoms. Findings are discussed in relation to clinical and technical limitations and future improvements.
All Science Journal Classification (ASJC) codes
- Hardware and Architecture
- Computer Vision and Pattern Recognition
- Computer Science Applications