Deep learning for EEG-based depression detection: A systematic literature review of methods and interpretability techniques.

Abstract

Depression is a widespread mental illness in which there is a noticeable difference in the quality of life of people. Electroencephalography (EEG) seems to be a promising non-invasive measurement for depression detection based on brain activity patterns. Recently, various deep learning approaches have been widely utilized in EEG data, enhancing accuracy and automation in depression detection. However, the inherent nature of deep learning models as a black box poses a challenge for their implementation in clinical practice, and thus interpretability would be a crucial part. This systematic review of the literature discusses more than 100 peer-reviewed selected studies published between 2017 and 2026, focusing on EEG-based depression detection using deep learning and interpretability methods. The review addresses multiple kinds of models, like convolutional neural networks, long short-term memory networks, and hybrid architectures, as well as interpretability approaches like shapley additive explanations (SHAP), local interpretable model-agnostic explanation (LIME), and attention mechanisms. Deep learning models are mostly identified to efficiently capture spatio-temporal EEG features without proper development in the adoption of interpretability techniques. Further challenges, however, are found in dataset limitations, non-uniformity in preparation and evaluation conditions, and barriers toward eventual clinical integration. Thus, this review serves as a complete synthesis of the current scenario and future perspectives on data availability towards enhanced reliability, transparency, and applicability of AI-driven depression diagnostics.