Robust discriminant feature extraction for automatic depression recognition

The incidence of depression has recently increased significantly. However, the current manual diagnosis may delay real-time detection and early treatment. Therefore, an automatic and effective auxiliary diagnosis is urgent. For automatic depression recognition, this paper presents a novel feature extraction algorithm, namely, Robust Discriminant Non-negative Matrix Factorization (RDNMF), which is joint optimization of the measurement of l(2,1)-norm, within-class scatter distance and between-class scatter distance. Different from traditional Non-negative Matrix Factorization (NMF) that just decomposes one high dimension matrix into the product of two new low dimension matrices, i.e. basic matrix and coefficient matrix, our algorithm also considers the robustness and discriminant of these two matrices, which can enhance the representation capability of basic matrix and significantly improve classification performance compared to other comparative methods. In addition, we have designed an audio stimuli paradigm for the measurement of functional Near-Infrared Spectroscopy (fNIRS) in task-state experiment. Finally, under the negative audio stimuli, our algorithm has promising results with random forest classifier, that is, Accuracy of 96.4%, Specificity of 100%, Sensitivity of 95.0% and AUC of 93.5%, which are superior in comparison with comparative machine learning methods, and simultaneously have comparable potential to state-of-the-art neural networks. Moreover, results also show that recognition rate of depression is highest under negative audio stimuli, which makes it possible to extract prominent features with this algorithm for auxiliary diagnosis of depression.