A self-organized QTPat-based feature-engineering framework for subject-independent classification of English- and...

Our primary aim is to develop an automated feature-engineering framework for subject-independent classification of EEG recordings acquired from English-speaking and Turkish-speaking participant cohorts. Drawing inspiration from a quantum tunnel pattern called QTPat, we introduced a novel approach of dynamically generating directed graphs to extract textural features. For our experiments, we utilized an EEG dataset containing 1652 English sentence-EEG pairs and 1600 Turkish sentence-EEG pairs. These signals were recorded from 20 English-speaking and 20 Turkish subjects who were presented with standardized sentences in their respective native languages. During the experiments, the participants were instructed to carefully contemplate each sentence while wearing a mobile brain cap system that acquired 14-channel EEG signals. Our preprocessing pipeline involved subjecting the EEG signals to multiple filtered wavelet packet decompositions, using both 6 Daubechies and 6 symlet wavelet filters. Subsequently, we extracted textural and statistical features employing the QTPat model, along with 20 statistical moments. To enhance discriminative power, we applied neighborhood component analysis and the Chi2-function to select the most relevant features. The k-nearest neighbor classifier was then employed to calculate predictive vectors. To further refine the results, we introduced an iterative hard majority voting scheme, generating additional voted vectors. Finally, we utilized a greedy algorithm to select the best-channel-wise and overall model results. Using leave-one-subject-out cross-validation (LOSO CV), the proposed framework achieved 96.37% classification accuracy using four selected EEG channels. These findings demonstrate that the QTPat-based architecture can discriminate the two language-associated EEG cohorts under subject-independent evaluation. However, because language and participant population are coupled in the available dataset, the present results should not be interpreted as definitive evidence of language-specific neural decoding. Further studies using within-subject bilingual experimental designs are required to determine whether the identified EEG patterns specifically reflect language-related neural processing.