Major Depressive Disorder (MDD) is a major public health concern, and its clinical diagnosis is mostly based on subjective assessments necessitating objective, data-driven diagnostic tools. Electroencephalogram (EEG) signals provide a non-invasive way to study brain activity and capture neurophysiological changes related to depression. However, EEG-based analysis is challenging due to high dimensionality, inter-subject variability, and the non-stationary nature of the signals. In this work, we compare two feature learning pipelines for automated MDD detection using multichannel EEG data recorded under resting-state and task-based conditions. The first pipeline uses handcrafted features. Local Binary Pattern (LBP) descriptors are extracted independently from each EEG channel to capture local temporal patterns. Features from all channels are concatenated at the subject level and ranked using the Fisher Discriminant Ratio (FDR). Forward feature selection is then applied, followed by classification using a Support Vector Machine (SVM). This pipeline emphasizes simplicity and interpretability of the extracted features. The second pipeline adopts a deep transfer learning strategy. Pretrained convolutional neural networks (VGG-19) are used to extract latent representations from individual EEG channels.. This approach is intended to model complex, non-linear EEG patterns that may not be captured by handcrafted features, while addressing the limitations of a small dataset. Experiments conducted on a publicly available EEG dataset with 64 subjects (34 MDD and 30 healthy controls) show that the deep feature-based pipeline achieves higher accuracy (76%) compared to the handcrafted LBP-based pipeline (63%). The results show the benefit of pretrained deep representations for EEG-based depression detection. Future work will focus on combining the two pipelines for better latent representation, subject-independent evaluation, end-to-end CNN models for EEG, and multimodal fusion. We are also developing a CNN pretrained on large public EEG datasets to improve generalization to unseen data.