Improving EEG major depression disorder classification using FBSE coupled with domain adaptation method based machine learning algorithms

• Published in: Biomedical signal processing and control Vol. 85; p. 104923
• Main Authors: Mohammed, Hadeer, Diykh, Mohammed
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.08.2023
• Subjects: Domain adaptation; Fourier Bessel series expansion; LS-SVM; Segmentation; Statistical features; Student t-test; Wilcoxon test; Wilcoxon test; Student t-test; Fourier Bessel series expansion; LS-SVM; Segmentation; Domain adaptation; Statistical features
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2023.104923

•FBSE is used to extract features from original data.•Statistical metrics are applied to remove noisy features.•DA is employed to reduce the difference of feature distributions among subjects. Major depression disorder (MDD) has become the leading mental disorder worldwide. Medical reports have shown that people with depression exhibit abnormal wave patterns in EEG signals compared with the healthy subjects when they are exposed to positive and negative stimuli. In this paper, we proposed an intelligent MDD detection model based on Fourier-Bessel series expansion (FBSE) coupled with domain adaptation (DA). First, EEG signals are segmented into intervals and each segment is passed through FBSE. Two types of features, including statistical and nonlinear features are investigated and extracted from each FBSE coefficient to detect MDD. Student t-test and Wilcoxon test are employed to remove noisy and bad features that can compromise the performance of data-driven learners. Then, DA method named Independence Domain Adaptation was applied to reduce the difference of feature distributions among subjects. The selected features are sent to a least square support vector machine (LS-SVM), and other classifiers named SVM, k-nearest (KNN), ransom forest, Bagged ensemble, boosted ensemble, decision tree, gradient boosting and stacked ensemble for the comparison purpose. Our experiments are simulated by using publicly available dataset. The performance of the proposed model is evaluated in both subject dependence experiment by 10-fold cross validation, subject independence experiment by leave-one-subject-out cross-validation, and Confidence interval respectively. Results showed that the features reduction method can significantly improve the mean accuracy by 4.20. The proposed model is compared with previous studies and the results show that the proposed model outperforms the other methods.