The diagnosis of depression is a critical topic in the medical field. For years, the electroencephalogram (EEG) has been considered an objective and cost-effective detection tool. However, most studies on depression recognition models tend to extract information solely from the original temporal scale of EEG signals, ignoring the usage of coarse scales. This study aims to explore the feasibility of multiscale analysis for a depression recognition model and to research its temporal characteristics. Based on two types of multiscale entropy, this paper constructs a machine learning model using classifiers including LDA, LR, RBF-SVM, and KNN. The relation between the temporal scale and model performance was examined through mathematical analysis. The experimental results showed that the highest classification accuracy achieved was 96.42% with KNN at scale 3. Among various classifiers, scales 3 and 9 outperformed other scales. The model performance is correlated with the scale variation. Within a finite range, an optimal scale likely exists. The algorithm complexity is linearly related to the temporal scale. By accepting predictable computational costs, a stable improvement in model performance can be achieved. This multiscale analysis is practical in building and optimizing a depression recognition model. Further investigation of the relation between the temporal scale and model capabilities could advance the application of computer-assisted diagnosis.

The early diagnosis of depression is often impeded by the subjectivity inherent in traditional clinical assessments. To advance objective screening, this study proposes a lightweight neural network framework designed to discriminate between pathological depressive states and non-pathological transient negative emotions using EEG signals. Diverging from conventional methods that rely on single-domain features, we construct a comprehensive multi-domain feature space via Wavelet Packet Decomposition. Specifically, the framework integrates frequency (α/β power spectral density ratio), spatial (normalized α-asymmetry), and non-linear (Sample Entropy) attributes to capture the heterogeneous neurophysiological dynamics of depression. To effectively synthesize these diverse features, a multi-head additive attention mechanism is introduced. This mechanism empowers the model to adaptively recalibrate feature weights, thereby prioritizing the most discriminative patterns associated with the disorder. Experimental validation on the DEAP (negative emotion) and HUSM (major depressive disorder) datasets demonstrates that the proposed method achieves a classification accuracy of 92.2% and an F1-score of 93%. Comparative results indicate that our model significantly outperforms baseline SVM and standard deep learning approaches. Furthermore, the architecture exhibits high computational efficiency and rapid convergence, highlighting its potential as a deployable engine for real-time mental health monitoring in clinical scenarios.