[Paper Review] Deep Convolutional Architectures for EEG Classification: A Comparative Study with Temporal Augmentation and Confidence-Based Voting

Aryan Patodiya, Hubert Cecotti|arXiv (Cornell University)|Feb 25, 2026

EEG and Brain-Computer Interfaces0 citations

TL;DR

The paper compares 2D and 3D CNN architectures for ERP-based EEG classification, showing that a 3D CNN with temporal shift augmentation and test-time confidence-based voting outperforms 2D variants, with CSP aiding 2D models.

ABSTRACT

Electroencephalography (EEG) classification plays a key role in brain-computer interface (BCI) systems, yet it remains challenging due to the low signal-to-noise ratio, temporal variability of neural responses, and limited data availability. In this paper, we present a comparative study of deep learning architectures for classifying event-related potentials (ERPs) in EEG signals. The preprocessing pipeline includes bandpass filtering, spatial filtering, and normalization. We design and compare three main pipelines: a 2D convolutional neural network (CNN) using Common Spatial Pattern (CSP), a second 2D CNN trained directly on raw data for a fair comparison, and a 3D CNN that jointly models spatiotemporal representations. To address ERP latency variations, we introduce a temporal shift augmentation strategy during training. At inference time, we employ a confidence-based test-time voting mechanism to improve prediction stability across shifted trials. An experimental evaluation on a stratified five-fold cross-validation protocol demonstrates that while CSP provides a benefit to the 2D architecture, the proposed 3D CNN significantly outperforms both 2D variants in terms of AUC and balanced accuracy. These findings highlight the effectiveness of temporal-aware architectures and augmentation strategies for robust EEG signal classification.

Motivation & Objective

• Motivate robust single-trial ERP classification in EEG for BCI applications amidst low SNR and latency variability.
• Evaluate and compare multiple deep learning pipelines (2D CNN with CSP, 2D CNN without CSP, and 3D CNN) for ERP detection.
• Investigate temporal augmentation (temporal shift) and test-time confidence-based voting to mitigate ERP latency jitter.
• Quantify performance across participants and establish whether 3D architectures outperform 2D architectures in a fair comparison.

Proposed method

• Preprocess EEG data with bandpass filtering, surface Laplacian, and normalization; segment into 1000 ms epochs and downsample to 64 time points.
• Compare three pipelines: 2D CNN with CSP (6 spatial filters, 64 time points), 2D CNN on raw 32 channels, and 3D CNN on 5D volumetric input.
• For 3D CNN, map 32-channel EEG onto a 7x5 grid, interpolate missing positions, and apply temporal shift augmentation to create shifts {-2,-1,0,1,2}.
• Train with focal loss and class weighting, apply batch normalization, GELU activations, and dropout; use stratified 5-fold cross-validation and early stopping based on AUC.
• During testing for 3D CNN, perform confidence-based voting across shifted trials and select the prediction with the highest softmax confidence.

Experimental results

Research questions

• RQ1 Can 3D CNNs jointly model spatiotemporal ERP patterns more effectively than 2D CNNs for ERP classification?
• RQ2 Does incorporating CSP preprocessing improve 2D CNN performance compared to raw-channel inputs?
• RQ3 Does temporal shift augmentation and test-time confidence voting enhance robustness to ERP latency variability?
• RQ4 What is the comparative performance gain of 3D CNNs over the best-performing 2D CNN under fair comparison conditions?

Key findings

• 3D CNN with temporal shift augmentation and confidence-based voting achieved the highest performance, with mean AUC up to 0.994 across participants using GAP variant.
• 2D CNNs benefit consistently from CSP preprocessing, yielding higher mean AUC (0.866 with CSP vs. 0.857 without CSP) for the best 2D model.
• Among 3D CNN variants, the GAP version delivered the best results (mean AUC 0.994, std 0.003).
• Best 3D model with GAP significantly outperformed the best 2D model with CSP (AUC 0.866) and without CSP (AUC 0.857); paired t-test indicates p < 0.001.
• Subject-specific evaluation shows strong per-subject performance; authors note lack of cross-subject generalization tests and potential for future transfer learning and interpretability enhancements.