Machine learning based classification of imagined speech electroencephalogram data from the amplitude and phase spectrum of frequency domain EEG signal.

Abstract

Imagined speech classification involves decoding brain signals to recognize verbalized thoughts or intentions without actual speech production. This technology has significant implications for individuals with speech impairments, offering a means to communicate through neural signals. The prime objective of this work is to propose an innovative machine learning (ML) based classification methodology that combines electroencephalogram (EEG) data augmentation using a sliding window technique with statistical feature extraction from the amplitude and phase spectrum of frequency domain EEG segments. This work uses an EEG dataset recorded from a 64-channel device during the imagination of long words, short words, and vowels with 15 human subjects. First, the raw EEG data is filtered between 1 Hz and 100 Hz, then segmented using a sliding window-based data augmentation technique with a window size of 100 and 50% overlap. The Fourier Transform is applied to each windowed segment to compute the amplitude and phase spectrum of the signal at each frequency point. The next step is to extract 50 statistical features from the amplitude and phase spectrum of frequency domain segments. Out of these, the 25 most statistically significant features are selected by applying the Kruskal-Walli's test. The extracted feature vectors are classified using six different machine learning based classifiers named support vector machine (SVM), K nearest neighbor (KNN), Random Forest (RF), XGBoost, LightGBM, and CatBoost. The CatBoost classifier outperforms other machine learning classifiers by achieving the highest accuracy of 91.72 ± 1.52% for long words classification, 91.68 ± 1.54% for long versus short word classification, 88.05 ± 3.07% for short word classification, and 88.89 ± 1.97% for vowel classification. The performance of the proposed model is assessed using five performance evaluation metrics: accuracy, F1-score, precision, recall, and Cohen's kappa. Compared to the existing literature, this study has achieved a 5%-7% improvement with the CatBoost classifier and extracted feature matrix.