TWD-DepNet: a deep network enhanced by three-way decisions for EEG-based depression detection
Abstract The non-invasive capacity of electroencephalography (EEG) to record cerebral activity has made it a potential technique for depression identification. However, existing deep learning models often struggle with the inherent uncertainty of EEG signals and limited capacity for extracting discr...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-08-01
|
| Series: | Journal of King Saud University: Computer and Information Sciences |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s44443-025-00196-y |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract The non-invasive capacity of electroencephalography (EEG) to record cerebral activity has made it a potential technique for depression identification. However, existing deep learning models often struggle with the inherent uncertainty of EEG signals and limited capacity for extracting discriminative spatiotemporal features. To address these challenges, a novel deep learning framework TWD-DepNet is proposed, which is enhanced by Three-Way Decision (TWD) theory to enable robust EEG-based depression detection. The framework encompasses a complete pipeline from signal preprocessing to dynamic decision optimization. Specifically, ICA-based denoising, multi-band filtering, and adaptive segmentationare applied to obtain high-fidelity EEG representations. Then, a lightweight convolutional backbone (DepNet) with multi-scale convolution is designed, depthwise separable layers, and dynamic channel attention to capture rich spatiotemporal patterns efficiently. Critically, a TWD-enhanced uncertainty quantification module is introduced, where EEG samples are partitioned into positive, negative, and boundary regions via fuzzy clustering, explicitly modeling ambiguous cases. A confidence-weighted cross-entropy loss is further designed to impose greater penalties on uncertain samples, encouraging the model to focus on hard-to-classify cases. Finally, iterative training and validation strategies are employed to enhance model robustness and generalization. Experimental results on public depression EEG datasets demonstrate the effectiveness of our approach, achieving a classification accuracy of 94.66% with improved sensitivity (95.92%) and specificity (93.08%) compared to baseline models. Extensive ablation studies further validate the effectiveness of each module. |
|---|---|
| ISSN: | 1319-1578 2213-1248 |