Robust Multi-State EEG Cognitive Classification via Optimized Time-Domain Features and CatBoost

This study introduces a novel framework for classifying multi-state cognitive processes using electroencephalogram (EEG) signals. By integrating optimized time-domain feature extraction with ensemble learning techniques, the proposed method achieves exceptional accuracy in distinguishing eight distinct cognitive states. The preprocessing pipeline employs finite impulse response (FIR) bandpass filtering (0.5–45 Hz) and Independent Component Analysis (ICA) for artifact removal, while feature extraction leverages Hjorth parameters and statistical measures. A comparative analysis of classification algorithms reveals CatBoost as the top performer, achieving 93.4% accuracy, followed by Neural Network (91.3%), SVM (89.7%), and AdaBoost (88.9%). CatBoost excels in discriminating complex states with computational efficiency, processing times ranging from 18 ms (SVM) to 32 ms (CatBoost), supporting real-time applications. The framework demonstrates robustness under varying signal quality, maintaining >91% accuracy at 10 dB SNR. These advancements set new benchmarks for EEG-based cognitive monitoring, with implications for adaptive systems requiring real-time neural feedback.