A Study of recent classification algorithms and a novel approach for biosignal data classification

Analyzing and understanding human biosignals have been important research areas that have many practical applications in everyday life. For example, Brain Computer Interface is a research area that studies the connection between the human brain and external systems by processing and learning the brain signals called Electroenceplograhpy (EEG) signals. Similarly, various assistive robotics applications are being developed to interpret eye or muscle signals in humans in order to provide control inputs for external devices. The efficiency for all of these applications depends heavily on being able to process and classify human biosignals. Therefore many techniques from Signal Processing and Machine Learning fields are applied in order to understand human biosignals better and increase the efficiency and success of these applications. This thesis proposes a new classifier for biosignal data classification utilizing Particle Swarm Optimization Clustering and Radial Basis Function Networks (RBFN). The performance of the proposed classifier together with several variations in the technique is analyzed by utilizing comparisons with the state of the art classifiers such as Fuzzy Functions Support Vector Machines (FFSVM), Improved Fuzzy Functions Support Vector Machines (IFFSVM). These classifiers are implemented on the classification of same biological signals in order to evaluate the proposed technique. Several clustering algorithms, which are used in these classifiers, such as K-means, Fuzzy c-means, and Particle Swarm Optimization (PSO), are studied and compared with each other based on clustering abilities. The effects of the analyzed clustering algorithms in the performance of Radial Basis Functions Networks classifier are investigated. Strengths and weaknesses are analyzed on various standard and EEG datasets. Results show that the proposed classifier that combines PSO clustering with RBFN classifier can reach or exceed the performance of these state of the art classifiers. Finally, the proposed classification technique is