Performance comparison of feature extraction methods for electroencephalogram-based recognition of Balinese script

Abstract

Recognizing Balinese script from electroencephalogram (EEG) signals remains a challenging problem due to low signal amplitude, non-stationary dynamics, and significant inter-subject variability. Despite previous attempts, no single feature extraction method has been universally effective in addressing these limitations. To fill this gap, this study systematically evaluates five feature extraction techniques—differential entropy (DE), power spectral density (PSD), discrete wavelet transforms (DWT), Hjorth parameters, and statistical features—on the Balinese imagined spelling using electroencephalography (BISE) dataset, which contains EEG recordings specifically designed for Balinese script recognition. For classification, both artificial neural networks (ANN) and support vector machines (SVM) are applied, and their performance is validated across multiple experimental settings. Results demonstrate that DE consistently provides more stable and discriminative features than the other methods, achieving the highest classification accuracy when combined with ANN. These findings highlight the potential of DE-based approaches to advance electroencephalogram driven Balinese script recognition, offering a culturally significant contribution to brain-computer interface (BCI) research and supporting future applications in inclusive artificial intelligence, digital heritage preservation, and assistive technologies.