Proceedings of the Annual Meeting of the Cognitive Science Society

Major depressive disorder (MDD) involves heightened sensitivity to negative outcomes and altered avoidance learning. While deep learning has advanced EEG-based depression detection, most studies focus on resting-state signals rather than single-trial task feedback and its behavioral relevance. Using a public feedback-locked EEG dataset from a probabilistic learning paradigm, we decode correct versus incorrect feedback trials separately in MDD and healthy controls (HC) and propose DBNet, a dual-branch network combining time-domain modeling with learnable wavelet time-frequency representations. The time branch captures feedback-related ERP dynamics and trial-wise temporal variability, whereas the time-frequency branch learns task-relevant spectral components via adaptive wavelets. DBNet outperforms baselines, with higher decoding performance in MDD. Visualizations indicate that the model emphasizes FRN/P3 time windows and fronto-midline activity, and its learned wavelet spectra show a stronger theta-band emphasis in MDD, consistent with error-monitoring processes. Subject-level representations from DBNet further exhibit stronger coupling with test-phase NoGo accuracy in MDD.