How Does a Single EEG Channel Tell Us About Brain States in Brain-Computer Interfaces ?

Technical Summary: Single-Channel EEG for Brain State Classification in BCIs

Problem Statement
While Electroencephalography (EEG) is a cornerstone of Brain-Computer Interface (BCI) research due to its non-invasiveness and high temporal resolution, current high-performance systems typically rely on multi-channel setups covering the entire scalp. These configurations suffer from high costs, complex configurations, and reduced mobility, limiting their adoption in "real-world" and daily clinical applications. Although single-channel EEG offers a promising alternative for wearable technologies, significant challenges remain in effectively decoding complex neurological phenomena (such as mental arithmetic and motor imagery) from a single electrode without the spatial information provided by multi-channel arrays. Furthermore, existing methodologies often focus on training on all channels and testing on one, or extracting features for specific tasks, leaving the generalizability of models trained on a single channel to other channels largely unexplored.

Methodology
The study proposes a pipeline to classify cognitive tasks using single-channel EEG data, leveraging Convolutional Neural Networks (CNNs) with a focus on spectral-temporal features. The workflow involves three primary stages:

1. Data Acquisition and Preprocessing: The study utilized three distinct datasets involving healthy subjects performing Mental Arithmetic (MA) and Motor Imagery (MI) tasks.
   • Dataset 1 & 2: Focused on MA tasks with varying electrode counts (22 and 30 channels).
   • Dataset 3: Focused on MI tasks (rest, hand, elbow) with 64 channels.
   • Raw signals were preprocessed (band-pass filtering, artifact removal, re-referencing) and segmented into epochs.
2. Feature Extraction: Instead of manual feature engineering, the study employed time-frequency analysis. Event-Related Spectral Power (ERSP) was used to generate spectrogram images (fixed at 224 × 224 pixels) from the EEG epochs. This approach captures both frequency and temporal dynamics, treating the data as 2D images suitable for CNN processing.
3. Model Architecture and Training Strategies:
   • Architecture: A shallow 2D CNN was proposed, consisting of two blocks of Conv2D layers (50 filters, 5×5 kernel, ReLU activation) followed by batch normalization and max pooling. This design aims to balance feature extraction capability with low computational load.
   • Strategy A (Multi-channel Training, Single-channel Testing): The CNN was trained on data from all available channels to learn broad spatial patterns. It was then tested individually on each channel to evaluate the discriminative power of specific electrodes.
   • Strategy B (Single-channel Training, Single-channel Testing): The CNN was trained exclusively on data from one specific channel and then tested on all other channels. This approach assesses the model's ability to generalize learned patterns from a single source to diverse conditions and channels.
   • Evaluation: The models were evaluated using 10-fold cross-validation (splitting data into 70% training, 15% validation, 15% testing) across 15,840 to 93,000 trials. Metrics included Accuracy, Sensitivity, Specificity, and F1-score.

Key Results
The study demonstrated that single-channel EEG systems can achieve robust classification performance for both MA and MI tasks:

• Overall Performance: The shallow CNN significantly outperformed traditional methods (such as shrinkage Linear Discriminant Analysis and Support Vector Machines) and deeper architectures (like GoogLeNet and ShuffleNet) in terms of the accuracy-to-computational-load ratio.
  • Dataset 1 (MA): Achieved a median accuracy of 93.75%.
  • Dataset 2 (MA): Achieved a median accuracy of 91.57% (compared to 75.9% with sLDA).
  • Dataset 2 (MI): Achieved a median accuracy of 87.80% (compared to 65.6% with sLDA).
  • Dataset 3 (3-class MI): Achieved a median accuracy of 86.75% (compared to 68.68% with SVM).
• Channel-Specific Findings:
  • Strategy A Results: Frontal and fronto-central channels (e.g., F4, AFF1h, FCC3h) showed the highest accuracy for Mental Arithmetic tasks (up to 98.33%), while central and parietal regions (e.g., Pz, CCP3h) performed best for Motor Imagery (up to 96.55%). Temporal channels generally showed lower performance for these specific tasks.
  • Strategy B Results: The model demonstrated strong generalizability. For instance, in Dataset 1, training on channels like POO1, POO2, and P4 allowed the model to achieve 100% accuracy when tested on channels such as P7, P3, and C3. In Dataset 2, training on specific channels yielded accuracies up to 91.55% for MA and 73.45% for MI when tested on other channels.
• Model Efficiency: The shallow CNN provided a favorable trade-off, avoiding the heavy computational load of deeper networks (like LSTM or GoogLeNet) while maintaining superior accuracy.

Significance and Claims
The authors claim that this study contributes to the development of practical, low-cost, and user-friendly single-channel BCI systems. The significance of the work lies in:

1. Feasibility of Single-Channel BCIs: Demonstrating that complex brain states can be decoded using a single electrode, potentially eliminating the need for cumbersome multi-electrode headsets in real-world applications.
2. Robust Biomarkers: Identifying specific EEG channels (ROIs) that are most effective for specific tasks, which can guide the optimization of electrode placement for wearable devices.
3. Generalizability: Proving that models trained on single channels can generalize to other channels, suggesting that the underlying neural features are transferable across the scalp, which supports the adaptability of EEG-based diagnostics.
4. Computational Efficiency: Validating that shallow neural networks are sufficient for this task, making them suitable for deployment on portable, resource-constrained wearable hardware.

The paper concludes that while multi-channel systems offer high spatial resolution, single-channel approaches supported by deep learning frameworks offer a transformative potential for deploying BCIs in everyday settings, balancing system simplicity with reasonable classification accuracy.