Single-channel attention classification algorithm based on robust Kalman filtering and norm-constrained ELM.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Electroencephalography (EEG) signal classification is vital for brain-computer interfaces (BCIs).
• Noise and signal fluctuations in portable, single-channel EEG devices reduce classification accuracy.
• Existing methods struggle with real-time signal variability and interference.

Purpose of the Study:

• To develop a robust method for accurate EEG-based attention classification.
• To overcome limitations of noise and signal fluctuations in portable BCI applications.
• To enhance the generalization and performance of attention classification algorithms.

Main Methods:

• Integrated Discrete Wavelet Transformation (DWT) and Independent Component Analysis (ICA) for noise reduction.
• Employed a robust Kalman filter with convex optimization to preserve essential EEG components.
• Utilized a norm-constrained extreme learning machine (ELM) with L1/L2 regularization for improved classification.
• Validated the approach using data from the Schulte Grid paradigm, TGAM sensors, and public datasets.

Main Results:

• The robust Kalman filter achieved superior denoising, with average AUCs of 0.8167 (self-collected) and 0.8344 (public).
• Maximum AUCs reached 0.8678 (self-collected) and 0.8950 (public).
• The proposed method outperformed traditional Kalman filtering, LMS adaptive filtering, and TGAM's eSense algorithm in noise reduction and attention classification accuracy.

Conclusions:

• Combining advanced signal processing and machine learning significantly improves EEG-based attention classification robustness.
• The proposed method offers enhanced generalization for BCI applications.
• Future research should focus on larger, diverse participant groups and broader applications like mental health monitoring and neurofeedback.