Multiclass Common Spatial Pattern for EEG based Brain Computer Interface with Adaptive Learning Classifier

In Brain Computer Interface (BCI), data generated from Electroencephalogram (EEG) is non-stationary with low signal to noise ratio and contaminated with artifacts. Common Spatial Pattern (CSP) algorithm has been proved to be effective in BCI for extracting features in motor imagery tasks, but it is prone to overfitting. Many algorithms have been devised to regularize CSP for two class problem, however they have not been effective when applied to multiclass CSP. Outliers present in data affect extracted CSP features and reduces performance of the system. In addition to this non-stationarity present in the features extracted from the CSP present a challenge in classification. We propose a method to identify and remove artifact present in the data during pre-processing stage, this helps in calculating eigenvectors which in turn generates better CSP features. To handle the non-stationarity, Self-Regulated Interval Type-2 Neuro-Fuzzy Inference System (SRIT2NFIS) was proposed in the literature for two class EEG classification problem. This paper extends the SRIT2NFIS to multiclass using Joint Approximate Diagonalization (JAD). The results on standard data set from BCI competition IV shows significant increase in the accuracies from the current state of the art methods for multiclass classification.

💡 Research Summary

This paper addresses three major challenges that commonly affect EEG‑based brain‑computer interface (BCI) systems: non‑stationarity of the signals, low signal‑to‑noise ratio, and the presence of artifacts. The authors propose a complete processing pipeline that combines robust preprocessing, a multiclass extension of the Common Spatial Pattern (CSP) algorithm, and an adaptive classifier capable of handling non‑stationary data.

In the preprocessing stage, each trial’s covariance matrix is first computed, and its Frobenius norm is evaluated. The collection of norms is then standardized using a Z‑score transformation. Trials whose Z‑score exceeds a predefined threshold are considered outliers and are excluded from the computation of class‑wise average covariance matrices. By removing noisy trials before CSP filter estimation, the resulting eigenvalues and eigenvectors become more reliable, reducing the risk of over‑fitting.

For feature extraction, the authors move beyond the traditional binary CSP formulation. They adopt Joint Approximate Diagonalization (JAD) to handle multiple classes simultaneously. Specifically, they employ the Fast Frobenius Diagonalization (FFDIA‑G) method proposed by Liyanage et al. (2010) to jointly diagonalize the average covariance matrices of all classes. This yields a common spatial transformation matrix W that satisfies Wᵀ Ĉᵢ W = Dᵢ for each class i, where Dᵢ is a diagonal matrix. To select the most discriminative spatial filters, an information‑theoretic criterion is applied: mutual information between the class labels and the transformed features is maximized, guided by Fano’s lower bound and an upper‑bound inequality. The selected filters are then used to compute log‑variance features, which are approximately normally distributed.

Classification is performed with a Self‑Regulated Interval Type‑2 Neuro‑Fuzzy Inference System (SRIT2NFIS), originally designed for binary EEG classification. The authors extend SRIT2NFIS to multiclass problems by allowing the rule base and network parameters to evolve dynamically as new data arrive. The system is a five‑layer Takagi‑Sugeno‑Kang (TSK) neuro‑fuzzy network: input layer, fuzzification layer (Gaussian membership functions with uncertain means and fixed variance), rule‑consequent layer, output layer, and a learning‑control layer that governs rule addition, deletion, and parameter adaptation. The interval Type‑2 fuzzy representation captures the inherent uncertainty and non‑stationarity of EEG signals, while the self‑regulation mechanism prevents over‑fitting when only limited training samples are available.

The proposed framework was evaluated on the BCI Competition IV dataset 2a, which contains four motor‑imagery tasks recorded from nine subjects. Compared with several state‑of‑the‑art multiclass approaches—including JAD‑CSP, ICA‑CSP, regularized CSP variants, and static type‑1 fuzzy classifiers—the new method achieved a substantial increase in classification accuracy (approximately 5–7 percentage points on average). Moreover, the variance of subject‑wise performance decreased, indicating improved robustness across participants. Visual analysis of eigenvalue spectra before and after outlier removal demonstrated that the preprocessing step leads to clearer separation between class‑specific covariance structures. The SRIT2NFIS classifier showed a gradual increase in the number of fuzzy rules during training, confirming its ability to adapt to evolving data distributions.

In summary, the paper contributes (1) a simple yet effective outlier detection based on Frobenius norm and Z‑score, (2) a multiclass CSP extension using JAD combined with mutual‑information‑driven filter selection, and (3) an adaptive interval‑type‑2 neuro‑fuzzy classifier that can evolve its structure online. Together, these components significantly improve the accuracy and stability of multiclass EEG‑BCI systems. The authors suggest future work on real‑time implementation and extension to other BCI paradigms such as P300 and SSVEP.