Spectral Analysis of EEG Biosignals for the Classification of Auditory Stimuli

Calderón-Reyes J.E.; Gallegos-Acosta A.E.; Álvarez-Rodríguez F.J.; Maita-Mamani M. · Publicaciones de la Facultad de Educacion y Humanidades del Campus de Melilla (2026) · DOI: 10.30827/publicaciones.v56i1.36843

Resumen

Introduction: The analysis of Electroencephalography (EEG) biosignals based on spectral representations enables the identification of frequency-based patterns associated with cognitive and behavioral responses. Therefore, this study covers the classification of auditory stimuli using spectrograms derived from EEG recordings, with the objective of detecting readings that may reflect atypical or distress-related activity, thus contributing to the development of adaptive and accessibility-oriented software architectures capable of integrating personalized Artificial Intelligence (AI) models. Method: The methodological approach followed the Data Science (DS) process, encompassing acquisition, preprocessing, feature extraction, and classification. EEG data were selected from a dataset aligned with the project’s sensors configuration and subsequently subjected to digital filtering, normalization, and segmentation to ensure signal reliability. Spectrograms were generated using Short-Time Fourier Transform (STFT), normalized, and structured for model training. Eight Convolutional Neural Network (CNN) configurations incorporating Squeeze-and-Excitation (SE) blocks, early stopping, and hyperparameter optimization were implemented, of which a total of 149 models were trained and evaluated using accuracy, precision, recall, and F1 score. Results: Of the 149 trained models, 45 met the established performance criteria, and 30 were identified as distinct. The five best-performing models were selected for final evaluation, with the highest achieving 91.67% test accuracy. Comparative analysis demonstrated the relevance of preprocessing strategies, attention mechanisms, and subject-level validation in improving generalization and interpretability. The selected CNN architectures exhibited consistent performance across unseen subjects. Conclusions: The integration of spectral analysis with CNN-based classification proved effective for identifying auditory stimulus patterns in EEG biosignals. The findings support the feasibility of incorporating personalized AI models into accessibility-driven software architectures, enabling adaptive interaction based on EEG-derived user responses. This approach demonstrates potential applicability in clinical, educational, and assistive contexts.

Instituciones: Universidad Nacional de San Agustín de Arequipa