Desenvolvimento de métodos computacionais para predição de séries temporais de EEG

Abstract

This work presents the development of an open-source textual interface for the analysis, processing, and prediction of electroencephalogram (EEG) time series. The tool was created to automate routines for preprocessing, functional connectivity analysis, and training of predictive models based on artificial neural networks, specifically CNN-LSTM architectures, aimed at both predicting signal sequences and classifying brain frequency bands. The proposal aims to meet the demand for transparent and accessible computational methods for applications in neuromodulation, brain-computer interfaces, and clinical diagnostics, areas that traditionally rely on proprietary and non-transparent software. The tool was evaluated on three distinct datasets: a sustained attention task, restingstate EEG, and visual sensory stimulation, demonstrating satisfactory performance in predicting temporal trends and frequency classification, with validation accuracies above 84% in two of the three scenarios. The system has been made publicly available on GitHub, ensuring free access and enabling future expansions and adaptations by the scientific and developer communities.