Related Experiment Videos
Toward Robust EEG Classification Using Adaptive CNN-Transformer and Inception Architectures and Signal Level
Vikas Reddy Venkannagari1, Shivansh Sharma1, Parthan Olikkal1
1Department of Computer Science and Electrical Engineering, University of Maryland Baltimore County, Baltimore, MD 21250, USA.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces a new framework for electroencephalography (EEG) classification, improving accuracy in brain-state monitoring tasks like emotion recognition. The method enhances subject-specific results but faces challenges in generalizing across different individuals.
Area of Science:
- Neuroscience
- Machine Learning
- Signal Processing
Background:
- Non-invasive electroencephalography (EEG) is valuable for brain-state monitoring (e.g., emotion recognition, deception detection).
- Challenges in EEG classification include noise, non-stationarity, limited data, and inter-subject variability.
- Existing methods struggle with robust and generalizable EEG data analysis.
Purpose of the Study:
- To develop a sensor-density-aware framework for EEG classification.
- To improve EEG classification accuracy by adapting deep architectures to different sensor densities and augmenting data.
- To systematically compare subject-dependent and subject-independent classification performance.
Main Methods:
- Proposed a sensor-density-aware framework using different deep architectures for low- (5-channel) and high-channel (62-channel) EEG data.
- Implemented a physiologically constrained signal-level data augmentation procedure.
- Evaluated performance on the LieWaves (5-channel) and SEED (62-channel) datasets using subject-dependent and subject-independent metrics.
Main Results:
- Subject-dependent accuracy significantly improved on the 5-channel LieWaves dataset with augmentation (97.14%) compared to without (92.91%).
- On the 62-channel SEED dataset, augmentation yielded a minor, non-significant improvement (98.52% with vs. 98.44% without).
- Subject-independent performance remained low across both datasets (approx. 57-58%), highlighting generalization challenges.
Conclusions:
- The proposed framework effectively enhances subject-dependent EEG classification accuracy, particularly for low-density setups.
- Signal-level augmentation combined with sensor-density-aware architecture selection is a promising approach.
- Cross-subject generalization remains a significant hurdle for practical, widespread EEG applications.