EEG microstates classification empowered with optimum-path forest using different distance measurements

Abstract

Electroencephalography (EEG) is considered one of the most important tests for neurological disease diagnosis, whose signals comprise microstate information regarding the spatiotemporal characteristics of human brain activity. However, interpreting and identifying such signals denotes a complex and time-consuming activity, thus motivating the use of automated approaches like machine learning techniques in many studies. Recent research has presented several techniques for detecting neurological disorders through analyzing EEG microstates. However, this area has room for advancements and improvements, using distinct methods and more in-depth scrutiny. Therefore, this paper proposes a new method for EEG signal classification through microstate analysis for diagnosing neurological diseases using a graph-based algorithm, namely the Optimum-Path Forest (OPF) classifier. Experiments were conducted over features extracted from EEG microstates obtained from the Temple University Hospital Abnormal EEG Corpus (TUAB) and the Schizophrenia EEG datasets. Such features are further processed using principal component analysis, t-distributed stochastic neighbor embedding, and uniform manifold approximation and projection dimensionality reduction techniques to improve computational efficiency and increase the classifier’s performance. Furthermore, this work evaluates 44 distance measurements in OPF’s graph modeling context. Finally, the experimental results show that with a reduced set of features extracted from the microstates and an appropriate distance measure, it is possible to obtain accuracy values equivalent to 100% with low processing time compared to raw data. In addition, the $k$-nearest neighbors and support vector machine classifiers were used to compare the experimental results.