Analysis of Machine Learning Models Using Proposed EEG Vowel Dataset

Abstract

Electroencephalogram (EEG) signals are produced by neurons of human brain and contain frequencies and electrical properties. It is easy for a Brain to Computer Interface (BCI) system to record EEG signals by using non-invasive methods. Speech imagery (SI) can be used to convert speech imaging into text, researches done so far on SI has made use of multichannel devices. In this work, we propose EEG signal dataset for imagined a/e/i/o/u vowels collected from 5 participants using NeuroSky Mindwave Mobile2 single channel device. Decision Tree (DT), Random Forest (RF), Genetic Algorithm (GA) Machine Learning (ML) classifiers are trained with proposed dataset. For the proposed dataset, the average classification accuracy of DT is found lower in comparison to RF and GA. GA shows better performance for vowel e/o/u resulting accuracy of 80.8%, 82.36%, 81.8% for 70 − 30 data partition, 80.2%, 81.9%, 80.6% for 60 − 40 partition data and 79.8%, 81.12%, 78.36% for 50–50 data partition. Whereas RF shows improved classification accuracy for a/i which is 83.44%, 81.6% for 70 − 30 data partition, 82.2%, 81.2% for 60 − 40 data partition and 81.4%, 80.2% for 50–50 data partition. Some other performance parameters like min. value, max. value of accuracy, standard deviation, sensitivity, specificity, precision, F1 score, false positive rate and receiver operating characteristics are also evaluated and anal- ysed. Research has proven that brain functions remains normal in patients with vocal disorders. Completely disabled patients can be equipped with such technol- ogy as this may be one of the best way for them to have access over the essential day to day basic requirement.