Feature Extraction and Classification of EEG Signals in Brain Computer Interfaces

Sučelje mozga i računala (engl. Brain Computer Interface, BCI) osigurava komunikacijski kanal između čovjeka i računala. Komunikacija se temelji na prikupljanju i analizi signala generiranih u mozgu. BCI sustav zahtijeva od korisnika sposobnost generiranja određenih moždanih obrazaca koje potom može detektirati, dekodirati i pomoću njih upravljati cjelokupnim sustavom. Prema tome, BCI sustav je potpuno neovisan o bilo kakvoj vrsti pokreta korisnika. U ovom su radu analizirani signali EEG-a koji predstavljaju zamišljanje pokreta lijeve odnosno desne ruke te stanje kada ispitanik miruje. Korišteni su javno dostupni signali EEG-a, a to su BCI Competition 2008 Dataset IIIa. Analiza uključuje izdvajanje značajki temeljenih na algebri kvaternionima te klasifikaciju pomoću algoritama temeljenih na stablima odluke. Algoritmi koji su se koristili za klasifikaciju su stablo odluke, algoritam slučajnih šuma, AdaBoost algoritam te ExtraTrees algoritam. Iako primjena kvaterniona pruža elegantan način prikaza signala i efikasan način računanja, točnost koja je postignuta ovim algoritmom, koristeći ovaj skup podataka, je relativno niska te nedovoljna za primjene u stvarnom vremenu.Brain computer interface (BCI) provides communication channel between human and computer. Communication is based on acquisition and analysis of signals generated in brain. BCI system requires from user the ability of generating certain brain patterns which can be detected, decoded and used for managing the entire system. Thus, BCI system is completely independent of any users movement. In this paper EEG signals which represent hand motor imagery and state when subject rests were analyzed. Publicly available signal base was used and it was BCI Competition 2008 Dataset IIIa. Analysis includes feature extraction based on quaternion algebra and classification using algorithms based on decision trees. Classification algorithms that were employed are decision tree, random forest algorithm, AdaBoost algorithm and ExtraTrees algorithm. Despite of fact that use of quaternion algebra provides elegant way of representing signals and computationally is efficient, accuracy that is obtained using this dataset is relatively low and is not sufficient for real-time applications