Abstract
In this paper, we propose a novel Convolutional Neural Network (CNN) approach for the classification of raw dry-EEG signals without any data pre-processing. To illustrate the effectiveness of our approach, we utilise the Steady State Visual Evoked Potential (SSVEP) paradigm as our use case. SSVEP can be utilised to allow people with severe physical disabilities such as Complete Locked-In Syndrome or Amyotrophic Lateral Sclerosis to be aided via BCI applications, as it requires only the subject to fixate upon the sensory stimuli of interest. Here we utilise SSVEP flicker frequencies between 10 to 30 Hz, which we record as subject cortical waveforms via the dry-EEG headset. Our proposed end-to-end CNN allows us to automatically and accurately classify SSVEP stimulation directly from the dry-EEG waveforms. Our CNN architecture utilises a common SSVEP Convolutional Unit (SCU), comprising of a 1D convolutional layer, batch normalization and max pooling. Furthermore, we compare several deep learning neural network variants with our primary CNN architecture, in addition to traditional machine learning classification approaches. Experimental evaluation shows our CNN architecture to be significantly better than competing approaches, achieving a classification accuracy of 96% whilst demonstrating superior cross-subject performance and even being able to generalise well to unseen subjects whose data is entirely absent from the training process.
Introduction
Brain-Computer Interfaces (BCIs) enable direct communication between the brain and external devices, typically through Electroencephalography (EEG). While traditional "wet-EEG" systems provide high signal quality using conductive gels, they are impractical for daily use due to their cumbersome setup. "Dry-EEG" headsets offer a portable, gel-free alternative but suffer from higher skin-electrode impedance and noisier signals. Standard decoding practices involve intensive manual preprocessing, which is time-consuming and risks losing salient information. This research addresses these challenges by automating the classification pipeline, moving directly from raw, noisy dry-EEG data to accurate classification using deep learning.
Experimental Setup
Data was recorded using a Cognionics Quick-20 dry EEG headset from four subjects. Participants fixated on four flicker frequencies (10, 12, 15, and 30 Hz) presented on an LCD monitor. Recordings focused on seven channels over the parietal and occipital cortex, areas primary to visual processing. Subject 1 (S01) provided 100 trials per class, while other subjects provided 20 trials to test performance under data-limited conditions.
Methodology
The proposed architecture is an end-to-end 1D Convolutional Neural Network (CNN) designed for SSVEP signals. The core component is the SSVEP Convolutional Unit (SCU), which integrates a 1D convolutional layer to capture temporal patterns, batch normalization to stabilize training against noise, and max pooling for dimensionality reduction. Unlike traditional methods, this model processes raw signals without prior manual filtering or feature engineering. To improve generalization across different subjects, a deeper variant with up to five stacked SCU blocks was developed to learn more abstract, subject-independent features.
Results and Discussion
The CNN demonstrated superior performance across several metrics. It achieved a peak accuracy of 96% for a single subject (S01), outperforming traditional Support Vector Machines (SVM) and significantly surpassing Recurrent Neural Network (RNN) variants like LSTM (57%). In cross-subject scenarios, the CNN maintained 89% accuracy, whereas SVM performance dropped to approximately 66%. Notably, for "unseen" subjects, the deeper CNN variant achieved 69% accuracy, a substantial improvement over the near-random performance of traditional baselines.
Beyond accuracy, the CNN proved computationally efficient. Training on an Nvidia GTX 1060 took only 4 minutes, compared to roughly 10 hours for a Vanilla RNN. These results suggest that deep learning can identify subtle, non-linear features within noisy dry-EEG data that are often discarded by human-designed filters.
Conclusion
While the study is limited by a small sample size of four subjects, it demonstrates the potential for high-accuracy, "plug-and-play" BCIs. Future work will focus on hybrid CNN-RNN architectures to capture both spatial and long-term temporal dependencies, alongside testing on larger, more diverse datasets to further enhance subject-independent generalization.
Summary updated