A novel EEG-based major depressive disorder detection framework with two-stage feature selection

Figures

• 

  Fig. 1. Proposed framework for MDD detection and severity prediction: Firstly, in the input module of the framework, the raw EEG signals are derived by Nerron-Spectrum-5 to obtain the EEG rhythm features, and the subjects are diagnosed and scored by a physician to get the HAMD-17 score. Then, in the data processing module, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta /\alpha$$\end{document}β/α features are extracted from the EEG rhythm features and are Z-score standardized together with the rhythm features to obtain standardized features. Subjects with HAMD-17 scores greater than 17 are labeled as MDD, and those with HAMD-17 scores less than or equal to 17 are labeled as non-MDD. Moreover, the HAMD-17 score directly served as an indicator of MDD severity assessment. Then, in the feature selection module, PCC carries out the first stage feature selection on standardized features, and RFE carries out the second stage feature selection on reserved features. Finally, LR and SVM are used as…

• 

  Fig. 2. MDD severity assessment performance on random selected 10 records

• 

  Fig. 3. Scatter chart of LNR and LNR-DF with two-stage feature selection

• 

  Fig. 4. The process of determining \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau$$\end{document}τ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\zeta$$\end{document}ζ according to LR, SVM, and LNR respectively in feature selection: a LR, b SVM, c LNR

• 

  Fig. 5. Weight importance of selected features shared by SVM-DF and LNR-DF