Abstract
Classifying EEG data is integral to the performance of Brain Computer Interfaces (BCI) and their applications. However, external noise often obstructs EEG data due to its biological nature and complex data collection process. Especially when dealing with classification tasks, standard EEG preprocessing approaches extract relevant events and features from the entire dataset. However, these approaches treat all relevant cognitive events equally and overlook the dynamic nature of the brain over time. In contrast, we are inspired by neuroscience studies to use a novel approach that integrates feature selection and time segmentation of EEG data. When tested on the EEGEyeNet dataset, our proposed method significantly increases the performance of Machine Learning classifiers while reducing their respective computational complexity.
Introduction: The Challenge of Non-Stationarity
Electroencephalography (EEG) data is notoriously difficult to process due to its inherent noise and low signal-to-noise ratio. A critical but often overlooked challenge is non-stationarity: the fact that the underlying statistical properties of neural signals change over time.
In the context of Brain-Computer Interfaces (BCI), as a user practices a task, their neural activity evolves. As shown in Figure 1, brain activation patterns shift and generally become more efficient as a user moves from a "Novice" to a "Skilled" state.
Most standard machine learning approaches treat an entire recording session as a static block of data. This overlooks the dynamic nature of the brain; a model trained on the "Novice" brain at the start of a session may struggle to accurately classify signals from the "Skilled" brain at the end.
Methodology: The "Two Heads" Approach
The proposed Two Heads method is a bio-inspired pipeline that treats different stages of the learning process as distinct data sources.
1. Data Segmentation
Instead of processing the whole dataset at once, the researchers split each subject’s 15-minute recording into two equal halves: 1H (1st Half) and 2H (2nd Half). This serves as a simplified model of the "Novice-Intermediate-Skilled" progression observed in neuroscience.
2. Feature Extraction and Dual Selection
Following the EEGEyeNet benchmark pipeline, features were extracted using band-pass filtering (8-13 Hz alpha band) and a Hilbert transform across 129 channels. The core innovation lies in applying Dual Feature Selection: filter-based feature selection (Univariate Feature Selection using ANOVA F-values) was applied to 1H and 2H separately. This allows the system to identify the specific neural "signatures" most relevant to each phase of the session.
3. Training and Fusion
Eight different machine learning classifiers—including XGBoost, KNN, and Random Forest—were trained on these separate pipelines. The final classification score is derived from a weighted average of the 1H and 2H results.
Results and Performance
The "Two Heads" method demonstrated significant improvements across all tested classifiers compared to State-of-the-Art (SOTA) benchmarks.
Accuracy Gains
As detailed in the performance data, the XGBoost classifier using the 1H approach achieved 98.6% accuracy, outperforming the SOTA CNN benchmark of 98.3%. Notably, the 1H (1st Half) data often yielded higher accuracy on its own than the 2H data or the combined 1H+2H score. This suggests that neural signals generated during active learning (the "Novice" phase) may be more distinct and easier to classify than the "quieter" signals of a skilled brain.
Computational Efficiency
Beyond accuracy, the method offers substantial efficiency gains. By narrowing the feature space for specific time segments, the models achieve a superior performance-to-computation ratio. For instance, the runtime for Gradient Boost was reduced from nearly 150 seconds (SOTA) to approximately 70 seconds using the 1H+2H approach, while AdaBoost's runtime was nearly halved.
Discussion and Future Work
The study demonstrates that simpler machine learning models, when combined with smart, bio-inspired feature selection, can outperform complex and computationally expensive Deep Learning architectures.
However, some limitations remain. The method has currently been validated only on the EEGEyeNet dataset. Future research should test this temporal segmentation on other major benchmarks like GigaDB or BCI Competition IV to ensure generalizability. Additionally, exploring how this approach might benefit deep learning architectures like Transformers or its application in real-time BCI systems for paralyzed patients remains a promising avenue for future study.
Summary updated