WS-BiLSTM-MA: wavelet scattering-based BiLSTM with mixed attention block for MDD recognition using multi-channel EEG signals

• Published in: IEEE transactions on instrumentation and measurement Vol. 74
• Main Authors: Zhang, Feng, Yang, Chunfeng, You, Linlin, Wang, Xiaojia, Yuan, Yonggui, Le Bouquin Jeannès, Régine, Shu, Huazhong, Xiang, Wentao
• Format: Journal Article
• Language: English
• Published: Institute of Electrical and Electronics Engineers 2024
• Subjects: Bioengineering; Life Sciences; Electrodes; Accuracy; Hospitals; Scattering; Feature extraction; Electroencephalography; Convolutional neural networks; Character recognition; Long short term memory; Standards
• ISSN: 0018-9456
• DOI: 10.1109/tim.2024.3502843

Major depressive disorder (MDD) recognition using multichannel electroencephalography (EEG) signals has profound clinical value with its richness and accessibility of temporal information, but such signals may suffer from nonstationarity and redundant characteristics. To cope with these problems, the wavelet scattering-based bidirectional long short-term memory network with mixed attention block (WS-BiLSTM-MA) network is proposed with three core modules: 1) wavelet scattering network is applied to build a wavelet scattering matrix which captures the deformation stability information from multichannel cleaned EEG signals; 2) bidirectional long short-term memory network (LSTM) is used to obtain a relevant scattering matrix which learns the potential temporal relationship from the wavelet scattering matrix; and 3) MA block has two self-attention blocks which gain the ability to further redistribute the weights of features and integrate the key comprehensive information from scattering coefficients (SCs) and relevant matrices, respectively. The performance of our proposed WS-BiLSTM-MA network is evaluated on two MDD datasets in both eyes closed (EC) and eyes open (EO) conditions with 19-channel EEG signals: the Hospital University Sains Malaysia (HUSM) dataset and Zhongda Hospital Southeast University (ZHSU) dataset. To ensure subject independence, the leave-one-subject-out validation (LOSO) experiment and blind test (BT) validation experiment are conducted. In leave-one-subject-out experiment, the WS-BILSTM-MA network, with the zeroth-order, first-order, and second-order SCs, presents high performance in terms of classification accuracy, precision, recall, and $F1$ -score whatever conditions in both the datasets. The BT experiment demonstrates the excellent ability of our framework with only zeroth-order and first-order SCs, where the classification accuracy, precision, recall and $F1$ -score can reach 98.80%, 99.90%, 99.71%, 99.81% and 99.81%, 99.78%, 99.17%, 99.47% in EC and EO conditions in the HUSM dataset, and 83.29%, 87.97%, 75.65%, 83.60% and 83.41%, 90.59%, 74.20%, 81.58% in EC and EO conditions in the ZHSU dataset, respectively. Compared with some state-of-the-art methods in the HUSM dataset, the WS-BILSTM-MA network can improve the performance of MDD recognition proving its clinical interest.