Robust semi-supervised extraction of information using functional near-infrared spectroscopy for diagnosing depression

Abstract

Depression has become one of the major psychological disorders faced by contemporary human beings, and the current depression diagnosis model, which is based on the doctor’s questioning as the main diagnostic basis, can no longer meet the requirements of early detection and treatment of depression. To this end, this paper proposes a novel feature extraction algorithm, Robust Semi-Supervised Information Extraction (RSSIE), which is a joint optimization process of the $l_{2,1}$-norm, the graph Laplace operator, and some data labels, different from the traditional Non-negative Matrix Factorization (NMF), or Conceptual Factorization (CF), which decomposes the original high-dimensional matrix into two low-dimensional matrices only, in contrast, our proposed algorithm takes into account the robustness of the features and the flow structure of the features, makes full use of the existing labeling information, enhances the ability of the base matrix to contribute to depression diagnosis, and significantly improves the classification accuracy compared to other relevant methods. In addition, we developed an audio stimulation paradigm for functional near-infrared spectroscopy (fNIRS) measurements in task-state experiments. Finally, our algorithm shows the best classification results for negative audio stimuli, i.e., accuracy (92.5%), specificity (93.3%), sensitivity (91.5%), and AUC (91.0%), which is superior to traditional machine learning algorithms and can be used as an effective feature extraction method for depression diagnosis.