en Computational Neuroscience Computational Neuroscience Original Original <div style="text-align: justify;"><span style="font-size:10pt"><span style="background:white"><span style="text-autospace:ideograph-numeric ideograph-other"><span style="vertical-align:baseline"><span style="font-family:&quot;Times New Roman&quot;,serif"><span style="border:none windowtext 1.0pt; font-size:12.0pt; padding:0in"><span style="color:black">Major depressive disorder (MDD) is a common mental illness that is hard to diagnose accurately because its symptoms are subjective and often overlap with other conditions. In this study, we present a deep learning framework based on electroencephalography (EEG) signals to automatically and objectively detect MDD. We extracted statistical features (mean, standard deviation, variance, and root mean square) and spectral features (power spectral density using Welch&rsquo;s method) from 128-channel EEG recordings of 24 MDD patients and 29 healthy controls. An L1-regularized LinearSVC was used to select the most informative features by removing irrelevant ones. Then, a bidirectional long short-term memory (BiLSTM) network with a mixed attention mechanism captured temporal dependencies and emphasized important brain channels. We evaluated the model using leave-one-subject-out (LOSO) cross validation, which is a strict way to test generalization. Our method achieved an accuracy of 92.45% and an area under the curve (AUC) of 0.980, outperforming traditional classifiers such as SVM, KNN, CNN, LSTM and XGBoost. After Holm correction for nine multiple comparisons, two EEG features remained statistically significant (adjust p < 0.05), indicating robust group differences at the single-feature level. Furthermore, their multivariate pattern within the BiLSTM model also showed excellent discriminative power, confirming that MDD is associated with both isolated and distributed network-level alterations. In addition, the attention mechanism provides interpretability by highlighting which EEG channels and times segments contributed most to each prediction, offering clinicians insights into the model&rsquo;s decision-making process. These findings provide proof-of-concept evidence that EEG-based deep learning may serve as a decision-support tool for MDD assessment, although larger multi-center studies are required before clinical application.</span></span></span></span></span></span></span></div> Bidirectional LSTM (BiLSTM), EEG, depression detection, feature selection, deep learning, MDD classification, attention mechanism. 0 0 http://bcn.iums.ac.ir/browse.php?a_code=A-10-8675-1&slc_lang=en&sid=1 Meisam Shafaee m.shafaee@hut.ac.ir 13700319475328460059199 13700319475328460059199 Yes Hamedan University of Technology, Hamedan, Iran Sajad Khoshrah 13700319475328460059200 13700319475328460059200 No Hamedan University of Technology, Hamedan, Iran