Artificial immune system and particle swarm optimization for electroencephalogram based epileptic seizure classification

Abstract

Automated analysis of brain activity from electroencephalogram (EEG) has indispensable applications in many fields such as epilepsy research. This research has studied the abilities of negative selection and clonal selection in artificial immune system (AIS) and particle swarm optimization (PSO) to produce different reliable and efficient methods for EEG-based epileptic seizure recognition which have not yet been explored. Initially, an optimization-based classification model was proposed to describe an individual use of clonal selection and PSO to build nearest centroid classifier for EEG signals. Next, two hybrid optimization-based negative selection models were developed to investigate the integration of the AIS-based techniques and negative selection with PSO from the perspective of classification and detection. In these models, a set of detectors was created by negative selection as self-tolerant and their quality was improved towards non-self using clonal selection or PSO. The models included a mechanism to maintain the diversity and generality among the detectors. The detectors were produced in the classification model for each class, while the detection model generated the detectors only for the abnormal class. These hybrid models differ from each other in hybridization configuration, solution representation and objective function. The three proposed models were abstracted into innovative methods by applying clonal selection and PSO for optimization, namely clonal selection classification algorithm (CSCA), particle swarm classification algorithm (PSCA), clonal negative selection classification algorithm (CNSCA), swarm negative selection classification algorithm (SNSCA), clonal negative selection detection algorithm (CNSDA) and swarm negative selection detection algorithm (SNSDA). These methods were evaluated on EEG data using common measures in medical diagnosis. The findings demonstrated that the methods can efficiently achieve a reliable recognition of epileptic activity in EEG signals. Although CNSCA gave the best performance, CNSDA and SNSDA are preferred due to their efficiency in time and space. A comparison with other methods in the literature showed the competitiveness of the proposed methods.