Abstract
In the field of brain-computer interface (BCI), selecting efficient and robust features is critical for artificial intelligence (AI)-assisted clinical diagnosis. This study introduces a stacked deep structure for feature selection, utilizing the stacked generalized principle where random projections are added to original features in a layer-by-layer manner. This process helps "open" the manifold structure existing in the original feature space, making it more linearly separable. Evaluated on the University of Bonn epilepsy EEG dataset, the model demonstrates superior performance in classification tasks compared to benchmarking models, proving that features selected by this new structure are more meaningful and helpful to classifiers.
Introduction
Electroencephalogram (EEG) signals are a gold standard for diagnosing neurological disorders like epilepsy. However, raw EEG data is notoriously high-dimensional and noisy, with relevant patterns often buried within complex, non-linear structures called manifolds. Standard feature selection methods—filters, wrappers, and embedded methods—often struggle when data is not linearly separable. This paper addresses this gap by creating a deep structure that progressively simplifies the feature space through iterative refinement.
Methodology
The authors introduce SDE-JS-Regression (Stacked Deep Embedded Regression with Joint Sparsity). The framework is built upon a base model using $L_{2,1}$-norm regression, which enforces "joint sparsity," ensuring that only a small subset of features consistently useful across all data samples is selected.
The core innovation is the Stacked Generalized Principle. Instead of a single pass, the model operates in layers. The output of one layer is multiplied by a random projection matrix and added back to the original features as a "random shift." This iterative process effectively "untangles" the data manifold, making the input for each subsequent layer more linearly separable and easier for AI classifiers to process.
Experimental Setup and Datasets
The study utilizes the University of Bonn epilepsy EEG dataset, composed of five groups (A–E), each containing 100 single-channel EEG segments. These groups represent various states, including healthy subjects (eyes open/closed), epileptic subjects during seizure-free intervals, and subjects during active seizure activity.
Researchers organized these signals into three specific classification tasks:
- Task 1: Healthy (Eye-opening) vs. Healthy (Eye-closing).
- Task 2: Seizure vs. Interseizure vs. Non-seizure (3-class problem).
- Task 3: Healthy vs. Epileptic (Binary classification).
Key Results and Insights
The SDE-JS-Regression model was compared against four benchmarks: the original E-JS-Regression, mRMR, RFE-SVM, and Relief. Across all tasks, the proposed stacked model consistently outperformed the benchmarks.
- Strategic Randomness: While randomness is often viewed as noise, here random projections are used to shift the data distribution enough to expose underlying patterns.
- Optimal Depth: Experimental results showed that 4–6 layers typically provide the best balance for EEG data. Adding too many layers eventually distorts the data distribution, leading to a performance drop.
- Classifier Agnostic: The features selected improved the performance of both Support Vector Machines (SVM) and Ridge Regression classifiers, indicating the features themselves were inherently more robust.
Conclusion
The stacked deep structure effectively leverages deep learning principles to handle the non-linearity of EEG signals while maintaining the interpretability of feature selection. While the current determination of hyperparameters like the number of layers relies on trial and error or cross-validation, future work will focus on developing adaptive methods to tune these parameters automatically during training.
Summary updated