Depression　is　a　common　disease,　the　traditional　diagnosis　of　depression　mainly　relies　on　doctors　combined　with　self-Assessment　scale　for　clinical　diagnosis,　which　is　not　only　affected　by　subjective　factors,　but　also　time-consuming　and　laborious,　so　it　is　prone　to　problems　such　as　untimely　diagnosis　and　low　accuracy　rate.　An　essential　objective　foundation　for　the　diagnosis　of　depression　is　provided　by　the　use　of　EEG　data.　Many　deep　learning-based　depression　detection　models　have　emerged　in　recent　years.　Examples　of　these　models　include　the　conventional　convolutional　neural　network　(CNN)　and　the　long　and　short　term　memory　network　(LSTM).　However,　due　to　the　differences　in　the　setting　of　their　network　parameters　or　the　selection　of　their　features,　problems　such　as　overfitting　of　the　network　model　and　low　accuracy　of　the　model　prediction　may　occur.　Consequently,　this　research　suggests　a　very　robust　hybrid　neural　network　for　depression　detection　method　in　order　to　overcome　the　aforementioned　issues.　First,　features　may　be　extracted　from　the　input　EEG　signal　by　merging　a　CNN　with　spatially　focused　feature　extraction　capability　with　a　bidirectional　long　and　short　term　memory　(BiLSTM)　with　holistic　feature　extraction　capabilities　in　past　and　future　time.　Then,　the　process　for　paying　attention　is　presented　to　filter　out　the　more　important　features,　followed　by　softmax　classification.　Experiments　show　that　the　Att-1D-CNN-BiLSTM　model　proposed　in　this　paper　performs　well　on　both　the　public　dataset　MODMA　and　the　self-harvested　dataset,　and　outperforms　the　traditional　depression　detection　model　with　good　generalization　ability.　©　2024　IEEE.