Machine Learning Based Comparative Analysis of One, Four, and Fourteen Channel EEG Vowel Datasets

Brain-Computer Interfaces (BCIs) have substantially enhanced the interaction between humans and computers. Particularly, benefiting individuals with speech impairments who can use the BCIs for communication, underscoring the importance of this research domain. In this study, we explore the classification of imagined vowels through the decoding of neural signals from the brain using Electroencephalography (EEG) technology with lower channel EEG devices and then compare the obtained results. A new EEG dataset is formed specifically for this research, utilizing a single, four and fourteen- channel non-invasive EEG device that records brain signals while a subject imagines a vowel. Artificial Neural Network (ANN), Genetic Programming (GP), Long Short-Term Memory (LSTM) and Random Forest (RF) machine learning (ML) algorithms were utilized to classify the generated vowel datasets. Confusion matrix-based results were obtained for all models. Comparison of results demonstrated that LSTM produces the highest classification accuracy of 93.8% for fourteen channels and 92.2% for four channel datasets whereas, for single channel dataset Random Forest classifier gives the best results with an accuracy of 83.44%. Other performance parameters like sensitivity, precision, specificity, F1 score were also compared to find the most suitable classifier for lower channel datasets among the four classifiers under study. The whole idea of this research lies in finding the utility of classifiers for datasets of devices with lesser electrodes. As lower channel devices can be conveniently used by people under medical trails and may one day replace bulky devices.